Texture segregation in the cat: A parametric study

Vision Res. Vol. 32, No. 2, pp. 305-322, 1992 Printed in Great Britain. All rights reserved

0042~6989/92 $5.00 + 0.00 Copyright 0 1992 Pergamon Press plc

Texture Segregation in the Cat: a Parametric Study PETER

DE WEERD,*

Received 25 Ocrober

1990;

ERIK VANDENBUSSCHE,*

GUY A. ORBAN*

in r~ised~or~ 4 My 1991

We have investigated how d@erent texture parameters a#ect texture segregation in the cat, and which strategies cats use to solve the segregation task. Five cats were presented with stimuli consisting of two adjacent panels. One side contained a square area of a p~rti&u~~rtexture embedded in a deferent background texture; the other side waggled with only the background texture. The animal’s task was to detect at which side the texture dtrerence was presented. Sensitivity for the texture dtflerence was assessed by making one aspect of the texture (in most instances the size of the texture elements) dependent upon performance by means of a staircase procedure. Among the most prominent parametric eflects are those of density and element position randomization. In general, segregation was optimumat intermediate densities and deteriorated at htrger and smaller densities. Element position randomization caused a slight but systematic decrease in segregation performance. Furthermore, we found texture elements at the border between different textures to be of primary importance for segregation. Which strategy the animals used for solving the segregation task depended upon the presence of random figure/background reversals in subsequent stimulus presentations during training. The animals learned to detect texture dl~eren~es I_ these reversals were present, and without reversals, they learned to identify the particular texture in the target square. Interestingly, parameter dependencies of segregation did not depend upon the detection strategy used. We have speculated that the two dtflerent strategies used by the cats to solve the segregation tasks are related to direrent hierarchical levels of texture segregation which can be traced back to direrent stages of texture processing in human models of segregation performance. Cat

Texture segregation

Figure/background

reversal

INTRODUCTION

In humans, it is well documented that texture differences can be a powerful cue for surface segregation (Beck, 1982; Julesz, 1981, 1984; Nothdurft, 1985b; Treisman & Gelade, 1980). When texture areas consist of texture elements which differ by a single quality, such as brightness, color, size or orientation, the border between these areas is highly salient to the human observer. It has been proposed that texture segregation is based upon local properties (Bergen & Julesz, 1983; Julesz, 1980, 1981, 1984, 1986) and that segregation is related to differential activation of some type of local detectors coding these local properties (Beck, 1983; Caelli & Julesz, 1978; Caelli, Julesz & Gilbert, 1978; Julesz & Bergen, 1983; Turner, 1986). Hence, the processes underlying texture segregation can be subdivided into a first stage involving initial, local filtering followed by a second stage of difference calculation and boundary extraction.

*Laboratorium voor Neuro- en Psychofysiologie, Katholieke Universiteit te Leuven, Campus Gasthuisberg, Herestraat, B-3000 Leuven, Belgium.

Threshold

Texture segregation in the cat has been first investigated by Wilkinson (1986). In that study, cats were shown to be capable of detecting a target square composed of a texture differing from the surrounding texture. Interestingly, after replacing the texture defining the target by the background texture and vice versa, more than half of the cats showed significant difficulties in segregating the new stimulus. This suggests that cats can use two strategies to solve a segregation task. It seems that performance in one group of cats relied on the presence of a particular spatial arrangement of two textures (identification strategy), whereas in the other group perfo~ance was based on the presence of texture differences (differencing strategy). In addition, Wilkinson (1986) limited herself to the mere demonstration that cats are able to solve the segregation task, and did not investigate how segregation performance was related to parameters of the texture stimulus. The demonstration of texture segregation in the cat would be more convincing if this capacity could be shown to be at least qualitatively similar to human performance for a number of parameter manipulations. Therefore, the aim of the present study was to perform a parametric study

of texture 305

segregation

in the cat and

to compare

its

306

PETER

DE WEERD

performance with existing data about human performance, in order to demonstrate that the cat can be used as an animal model of human texture segregation. The demonstration of qualitatively similar segregation was important to us since we intend to use cats trained in texture segregation as subjects in lesion experiments. In addition, we wished to control the strategy the animals used to solve the segregation task and to assess the effect of the strategy used on parametric dependencies of texture segregation. We have employed the same basic task as Wilkinson (1986) did, and we have also used two of the same textures used by this author: the line texture stimulus and the dot/annulus texture stimulus. In the line texture stimulus, the target consisted of line segments surrounded by a background of differently oriented line segments. In the dot/annulus texture stimulus on the other hand, the target consisted of dots in a ground of annuli (or vice versa). The line texture stimulus is an interesting one, since experimental variations of parameters such as the difference in orientation between line elements in target and background, variations of line element length at constant width, and manipulations of the degree of orientation randomization of line elements have already been carried out in humans by Nothdurft (see Discussion). Hence, by using similar parameter manipulations we were able to compare human and feline segregation performance. Using the line texture stimulus, we also investigated the following, more specific questions, most of them also addressed in humans by Nothdurft. Firstly, we determined the efficiency of texture segregation as a function of two global stimulus variables: element density and element position randomization. Secondly, we investigated whether the effect of these two global variables interacted with target area. Thirdly, the effect of reducing the target to its border was studied, in order to verify Nothdurft’s hypothesis that the sudden change at the border between target and background (structure gradient) is the most important determinant of segregation. Finally, we investigated whether orientation processing of line elements during texture segregation is as efficient as during orientation discrimination of isolated line elements. In humans, the orientation difference required for segregation can be five times larger than differential orientation threshold determined with isolated line elements (see Discussion). This suggests that the output of initial orientation filtering is used differently in segregation tasks and classic orientation discrimination tasks. With respect to the dot/annulus texture stimulus, we investigated the effect of the size of the elements, element density, element position randomization and target area on segregation performance. We were interested whether these four parameters affected segregation performance in a way similar to line texture stimuli, despite the fact that orientation cannot be used as a cue in the segregation of dot/annulus textures. If the segregation of dot/annulus texture stimuli were to follow different rules compared to line texture stimuli, this would suggest that certain aspects of texture

ei al.

processing differ between the two texture stimuli used (see Discussion). In order to confirm our inference from Wilkinson’s (1986) study that cats can use two different strategies to solve the segregation task, we have compared texture segregation of line element textures obtained under the following two conditions. In the standard condition used in all experiments except one, target and background textures were randomly exchanged from trial to trial to force the animals to use the texture difference as a cue in the segregation task. On the other hand, segregation performance was also assessed in a condition in which the arrangement of figure and background textures was identical during the testing, and in which it could be established that the cats used an identification strategy instead of extracting the texture difference. In each of these two conditions, the effect of element length variations on segregation performance was assessed under different conditions of element density and element position randomization to evaluate the effect of the detection strategy used on segregation performance.

METHODS

Subjects

Five adult cats were used in the experiments: Jules (cat 57), Harry (cat 59), Poekie (cat 63), Mousty (cat 64), and Lollipop {cat 65). During training and testing, body weight was maintained at approximately 90% of the free feeding body weight. Testing apparatus and behavioral task

The cats were trained and tested in an apparatus designed after Berkley (1970). Briefly, this apparatus consist of a box, with a Plexiglas chamber mounted on the front panel of the box. The front panel of the Plexiglas chamber is divided into two transparent response keys, which the cat has to push in order to make a response. During the training and testing sessions, the animal was enclosed in the box from which it could thrust its head into the Plexiglas chamber. From this chamber it was able to view the stimuli through the response keys. The stimuli were presented on a screen positioned 25cm in front of the response keys so that viewing distance from the screen was approx. 28.5 cm. Behind one of the two response keys, the target texture was presented, surrounded by a background texture. The target texture occupied a square region with standard dimensions of 8.00 by 8.00deg. Behind the other key, only the background texture was visible. The animal’s task was to detect the stimulus side on which the target square was presented, and therefore we consider our segregation task to be a detection task. Thus, the presence of a target texture embedded in a background texture was the positive stimulus (S+) whereas the negative stimulus (S -) consisted of only the background texture (Fig. 1). Each detection trial started with the presentation of a texture stimulus. A separator inserted between the two

TEXTURE

SEGREGATION

response keys prevented the animal from seeing the left and right part of the stimuli simultaneously. During the first 0.35 set of stimulus presentation, all responses were ignored. This time period is called the Response Delay Period (RDP). The RDP was introduced to extinguish random responses immediately after stimulus onset. After expiration of the RDP, the animal was rewarded for pushing the response key behind which the target (S-t) appeared. Food rewards (pureed beef) were given through an opening in the bottom of the Plexiglas front chamber, just below the response keys. Incorrect responses were left unpunished. The S+ was presented behind the left and right response keys in a balanced, p~udorandom order, which allows efficient detection of any left- or right-key response preferences, and which provides objective criteria for compensating for such response preferences (De Weerd, Vandenbussche & Orban, 1990b). After the animal’s response, the stimulus disappeared, and the animal had to wait for a 4.00~set period until the

IN THE CAT

307

following stimulus was presented. This waiting period is the Inter-Trial Interval (ITI). Stimuli Stimulation apparatus. The stimuli were computergenerated images presented on a high-resolution, monochrome ATRIS screen, with a 75 Hz refresh rate. This screen is characterized by a 18 by 36 cm surface containing 720 by 1440 pixels respectively. Hence, both the S+ and the S- were generated in a 720 by 720 pixel field. Each individual pixel was either black (0.08 cd/m’) or white (24.02 cd/m’), which corresponds to a contrast between white and black of 2.48 [log(AZ/Z)]. Before a testing or training session, the texture stimuli (images) were generated and stored on harddisk. During the session, the images were retrieved from the harddisk during the ITI. Texture generation:

two types of texture elements.

As defined in the introduction, we used two different types of texture stimuli. Figure l(A) and (C) show

FIGURE 1. Representative sample of texture stimuli used in the different experiments. (A, B) A dot/annulus and a line texture stimulus at a 4.Wdeg element spacing respectively. (C, D) A dot/annulus and a line texture stimulus, but this time at a 20%deg element-spacing. In (A-D), element-size is slightly smaller than the element spacing. (E) Line texture stimulus at a 4.Wdeg element spacing in which the orientation difference between target and background line elements has been reduced. (F) Line texture stimulus at a 2.00-deg spacing in which the elements are randomized in orientation. The texture elements in the left hand column of stimuli are randomized in position (A, C, E), whereas regular texture stimuli are presented in the right hand column (B, D, F).

308

PETER DE WEERD et al.

representative examples of dot/annulus texture stimuli, and Fig l(B, D, E, F) show representative examples of line texture stimuli. In all stimuli used, texture elements were white on a black background. Within a given figure or background texture, all texture elements possessed identical properties: in a particular line texture, all elements were characterized by an identical length and width (standard width is 0.20deg), and in most cases also by an identical orientation. Likewise, in a dot texture, all dot texture elements were characterized by an identical diameter, and in the annulus texture all annuli were identical. Furthermore, care was taken that there was no significant luminance difference between target and background [see Appendix (I)]. Texture generation: global characteristics of a texture. During texture generation, the texture elements were dispersed over the image following a predefined grid. At each point of the grid, a texture element was generated. In one condition, referred to as the regular texture condition, the center of each texture element corresponded exactly to a grid point. Here, the texture elements were organized in regular rows and columns [Fig. l(B, D, F)]. In the other condition, referred to as the randomj~ed texture condition, texture elements were displaced from the grid points by a distance equal to a maximum of 50% of the spacing between neighboring grid points, with the direction and distance of this displacement randomly distributed [e.g. Fig. l(A, C, E)]. A procedure was followed which minimized overlap between neighboring elements [see Appendix (2>]. The spacing between neighboring grid points will be referred to as the element spacing. Length of the line elements will often be expressed relative to the element spacing, and this ratio is referred to as the relative length. Expressing the outer diameter of an annulus relative to the element spacing results in the relative diameter. Texture generatjon: manipulating the texture dl~erence. In order to assess the sensitivity of our subjects to a particular texture difference, series of stimuli were generated in which the dissimilarity of the textures defining target and background was progressively decreased. In dot/annulus texture stimuli, this was accomplished by reducing the diameter of the annuli and dots, keeping the area of the two texture elements equal. In line texture stimuli, any one of the following three manipulations could be carried out: a reduction of the length of the line elements, a reduction of the orientation difference between figure and background elements [Fig. l(E)] or an increase of the degree of orientation randomization of the line elements [Fig. l(F), see also Appendix (3)]. The manipulation of these parameters was done in a proportional way [see Appendix (3)]. Note that while manipulating a particular parameter of a texture, all other parameters of the texture stimulus were held constant. Where length of the texture elements varied, the orientation difference was maximal and orientation randomization was absent. Variations of o~entation difference were done at a relative length of 0.95 and without orientation randomization. Finally, orientation randomization was carried out at the maximal

orientation difference and at a relative length of 0.95. For each discrete step in the manipulation of a single parameter, four stimuli were generated: firstly, two stimuli were required to present the target behind the left or the right response key. Secondly, for each of these two stimuli, two versions were created in which the textures of figure and background were reversed. In addition, in each of the four resulting stimuli, the position of the target square was randomly chosen (Fig. 1). One of these four stimuli was randomly selected for each presentation of a particular value of the manipulated variable, and therefore, the randomization of target position and the random reversal of figure and background textures applied to all stimulus presentations (except if mentioned otherwise). These two randomizations are distinct from element position randomization, which was only applied in randomized texture conditions and not in regular texture conditions. Each parameter was manipulated at different values of element spacing [Fig. 3(C)]. We used 5 standard element spacings of 4.00, 2.00, 1.OO,0.50 or 0.3 1 deg. The same spacings were used in both regular and randomized texture conditions. Decreasing element spacing corresponds to increasing density. The densities corresponding to the 5 standard spacings used, are 0.06, 0.25, 1.00, 4.00 and 10.56 elements per square visual degree, respectively. Training procedures Initial training was carried out using a fading procedure, using stimuli presented from slides. During this initial training phase and subsequent testing, the target consisted of horizontal line elements surrounded by a background of vertical elements, and this texture arrangement remained fixed in all trials (see Results for more information). The use of the Differential Exposure Methods (DEM, De Weerd, Vandenbussche & Orban, 1990a) was required to train the animals to detect the target square once stimulus pre~ntation included random exchange of target- and background textures between trials (carried out in the ATRIS set-up). During training, we have continued using the stimuli made of horizontal and vertical line elements. In the three cats (57, 59, 63) trained with the line texture stimulus, the DEM was used to reduce the length of the elements, and in the two cats (64, 65) which were trained with the dot/annulus texture stimulus, the DEM was used to decrease the diameter of the dots and annuli. The DEM training started at a relative length or annulus diameter of 0.95. In a first block of 20 trials, the S- was occluded almost immediately after stimulus onset (0.05 set after expiration of the RDP). This was accomplished by instantaneously erasing S - , after which only the target (S+) remained in view, entailing the animal to quickly learn to respond to the target. After reaching a 75% correct performance, a second 20-trial block was administered in which the moment of occlusion was increasingly delayed in a performancerelated way (see De Weerd et al., 1990a). Since the

TEXTURE SEGREGATION IN THE CAT

random exchange of figure and background textures applied to all training trials, the animal progressively acquired the ability to detect a texture &@ence. After again reaching the 75% correct criterion, a third 20-trial block was presented in which both S+ and Sremained in view until a response occurred. This time, reaching criterion was followed by a reduction in length or diameter, after which the procedure just described was restarted (for further details, see De Weerd ef al., 199Oa). One training session consisted of 15 blocks of 20 trials. Unless the animals showed severe learning difficulties in the initial training phase, the DEM training was stopped after four sessions without improvement. The animals working with line texture stimuli were trained in four of the live standard element spacings (4.00, 2.00, 1.00 and 0.50 deg). Animals working with the dot/annulus texture stimulus were trained in all five standard spacings (including 0.31 deg). The complete DEM training in one condition of element spacing took 24 sessions on average (range 8-64 sessions). The order in which the different spacings were trained was randomized over cats. The duration required to complete training at the first element spacing was significantly longer (32-64 sessions) compared to training duration at subsequent element spacings (8-23 sessions). All cats were trained only for regular textures, in which the texture elements were arranged in rows and columns. The animals easily transferred from regular to randomized textures. Furthermore, the animals working with the line texture stimuli easily transferred to tasks in which the orientation difference between figure and background textures was manipulated and this was also true when the orientation variation of indi~dual line elements was introduced. In none of these cases was additional training given. Upon completion of the training and after initial testing by means of length thresholds, the cats were successfully transferred to stimuli made of left oblique and right oblique line textures. This transfer was necessary since it was found at this stage of the study that horizontal line elements were slightly brighter than verticals [see Appendix (4,5)]. Testing procedures

Detection thresholds were determined using a 73.5% correct Wetherill and Levitt (1965) staircase procedure. In such a staircase procedure, a single parameter of the texture stimulus was made dependent upon performance, just as described for the training. Notice that the distinctness of the texture difference with decreasing length, diameter or orientation difference and with increasing orientation randomization [see Appendix (311. The distinctness of the texture difference was reduced after three consecutive correct responses, or after two correct responses foilowed by an incorrect and a correct one. The texture difference was increased after an incorrect response, a correct response followed by an incorrect response, or after two correct responses followed by two incorrect ones. In a single session, four staircase thresholds were determined, each staircase lasting 75 trials. All staircase measurements started at threshold level as determined from previous sessions. The first 5

309

trials of each staircase did not contribute to threshold estimation. Since the staircase measurements started at threshold, all reversal points were included in threshold calculation. Because of the proportional adaptation of the independent variable, the threshold was calculated as a geometrical mean. Data collection and analysis

Four staircase thresholds were determined for each condition in each cat. Different conditions were administered in random order, except for the two conditions of element position randomization. For example, when element spacing was manipulated, the animals were tested during each session with a randomly chosen element spacing, but within a given session, two thresholds were determined using regular textures, followed by two other measurements using randomized textures, or vice versa. In some experiments, only randomized textures were used. In these experiments, the animals were tested daily at two different element spacings, two thresholds being devoted to each spacing. Data were analyzed using randomized factorial block analysis of variance (ANOVA; Kirk, 1968). A logarithmic transformation was used to achieve homogeneous variance. The ANOVAs were carried out upon the raw data, meaning that every single threshold determined in a session is included in the analysis. This was accomplished by the introduction of “replication” as an additional factor in the analysis. This factor was never significant, nor did it significantly interact with other factors. Reaction times were analyzed, but no relation was found with segregation performance, nor was there a relation with the number of texture elements in the stimulus or with the size of texture elements. These negative findings, as well as the fact that reaction times were long (1 set or more), suggest that they reflect mainly non-visual factors.

RESULTS In a first set of experiments, we determined length thresholds as a function of element spacing using the line texture stimulus (cats 57, 59 and 63) and annulus diameter thresholds using the dot/annulus texture stimulus (cats 64 and 65). In this experiment, the standard target square of 8.00 by 8.00deg was used. In two following experiments, length threshold and diameter thresholds were again determined for line- and dotlannulus texture stimuli, but the surface of the target square was manipulated in addition to the element spacing for both types of texture stimuli. Finally, we did a series of experiments involving the line texture stimulus only, using the standard target square. Firstly, the effect of width upon length thresholds was investigated for different element spacings. Secondly, orientation difference thresholds were determined as a function of element spacing. Thirdly, orientation randomization thresholds were assessed as a function of element spacing. In addition,

PETER

310

DE WEERD

these three experiments were repeated after reducing the target to the outer boundary of the target square. The experiments are described in a chronological order unless mentioned otherwise. Each figure presents data from a separate experiment, except for some figures in which data from previous experiments are repeated for comparison. Whenever this is the case, it will be explicitly mentioned. Size of texture elements at detection threshold as a function of element spacing In an initial experiment (Fig. 2), two cats were tested for a large number of element spacings, including the element spacings for which the animals were trained (4.00, 2.00, 1.OO, 0.50 and 0.31 deg) as well as intermediate element spacings for which the cats had received no prior training. The results obtained from cat 57 with line texture stimuli are shown in Fig. 2(A) and the results from cat 64 with dot/annulus texture stimuli in Fig. 2(B). Length of the line element [Fig. Z(A)] and diameter of the annulus [Fig. 2(B)] at the 73.5% detection level are plotted as a function of decreasing element spacing. Both length and diameter of the texture elements are referred to as element size. For each of the two testing conditions and in each cat, a curve fitting the size thresholds for the complete range of element spacings, would be U-shaped and tilted 45 deg, with a broad base at intermediate spacings (Fig. 2). Fu~hermore, note the systematic elevation of size thresholds in randomized texture conditions and the extension of segregation performance towards smaller element spacings for the dot/annulus texture stimulus compared to the line texture stimulus (Fig. 2). When only considering performance at intermediate spacings, size thresholds closely fit a proportional relation

A

T

CAT

57

ef al.

with element spacing. Indeed, after log-transformation of spacings and thresholds, the slope of the regression lines through the thresholds obtained at intermediate spacings was not significantly different from 1. Length thresholds obtained with line texture stimuli were proportional to element spacing for spacings ranging from 2.67 to 1.60 deg [Fig. 2(A)]. The proportional relationship between diameter threshold and spacing held for spacings varying between 2.00 and 0.50 deg in the regular texture condition and for spacings varying between 1.60 and 0.50 in the randomized texture condition [Fig. 2(B)]. The invariance of relative length thresholds with element spacing at intermediate spacings has also been described in human subjects (Nothdurft, 1985b). This indicates that in both species, the visual system scales element size for element spacing in order to reach a particular detection performance. This scaling mechanism must therefore be an essential mechanism in texture segregation. The breakdown of this scaling mechanism at extreme element spacings is a convenient indicator of decreasing sensitivity to the texture difference. Thus, size thresholds expressed relative to element spacing give a direct indication of the efficiency of texture segregation (relative length or relative diameter thresholds). In a second experiment, we replicated the data obtained in the initial experiment and extended our findings to all five cats. Only the standard element spacings were used in this experiment and in all subsequent ones. Relative length thresholds were determined as a function of element spacing in the regular texture and the randomized texture condition. A representative sample of stimuli used in this experiment is shown in Fig. 3(C). Data from cats 57, 59 and 63, as well as averaged

B

CAT

64

x

0.5i

O----n

I r, FIGURE 2. dot/annulus (cat 57). (B) lines refer to the data are

random.

I 2

0.25t

o- - - 0

random.

I 1 element

I spacing

(vrsual

I

I

2

1

I 0.5

031

degrees1

Absolute size thresholds plotted as a function of element spacing for line element texture stimuli (squares) and texture stimuli (circles). (A) Lengths allowing a 73.5% correct performance are plotted against element spacing Annulus diameter thresholds are plotted as a function of element spacing (cat 64). Open symbols and stippled the randomized texture condition; solid symbols and full lines refer to the regular texture conditions. Notice that presented in a log-log plot. The lines plotted through the data are determined by linear regression (see text for further explanation). Error bars represent standard deviations.

TEXTURE

SEGREGATION

311

IN THE CAT

thresholds obtained with regular textures. Finally, we data are shown in Fig. 3(A). All features of these data were unable to determine thresholds at the OSO-deg agree with the data presented above [Fig. 2(A)]. Firstly, element spacing. Notice that the range of optimal element the significant effect of element spacing upon detection performance was confbmed [F(2,46) = 38.32; P -cO.OOS].spacings is larger than that suggested by Fig. 3(A) [compare with Fig. 2(A)]. The sharp inflection in the Relative length thresholds were lowest at the intermediate curves is due to the large step between the different element spacing of 2.00 deg, and increased with smaller standard element spacings. and larger element spacings, an observation which Figure 3(B) shows detection thresholds of relative held for both regular and randomized texture stimuli. annulus diameters as a function of element spacing, Furthermore, in agreement with Fig. 2(A), thresholds obtained with dot/annulus texture stimuli in cats 64 and obtained with randomized textures were consistently 65. Again, the data of the present experiment confirm the elevated [P(1,46) = 19.73, P < O.OOOS} compared to A CAT

CAT

57

CAT

59

63

AVERAGE 12

12_

1.2

M n-

1.0

-cl

random

10

lO_

[_

regular -

0.8

0.6t 04 0.2

OOL

0.0 L I

I

J

4

2

1

CAT

CAT

61

1 L

f

,

2

1

AVERAGE

65

w

regular

O---O

random

06..

etement

spacmg

ELEMENT 4

degrecrs

h’lsual

SPACING 2

1

(visual

degrees)

1

FIGURE 3. (A) Relative length thresholds plotted as a function of element spacing for cats 57, 59, and 63. Data averaged over cats are shown in the right hand graph. (B) Rehttive annulus diameter thresholds are plotted as a function of element spacing for cat 64 and 65 separately. Again the average data are shown on the right. Open symbols and stippled lines refer to the randomized texture conditions, solid symbols and full lines refer to the regular texture conditions. Error bars represent standard deviations, except in the summary graphs where they represent standard errors. (C) Randomized (rand.) and regular line texture stimuli are shown at the three element spacings used. The lengths of the line segments in the stimulus reproductions correspond to the average length thresholds reached by the cats. YR 32/2-E

312

PETER DE WEERD et

data obtained in the first experiment [Fig, 2(B)]. Firstly, element spacing significantly affected the relative size thresholds [F(3,31) = 13.92, P < O.OOl]. Relative annulus diameter thresholds were lowest at intermediate element spacings, but the range of element spacings corresponding to minimum thresholds was wider than for line texture stimuli, in agreement with the data shown in Fig. 2. Furthermore, the present experiment confirmed that relative diameter thresholds could be measured at element spacings for which relative length thresholds could not be determined. Finally, as in the line texture stimulus, randomized textures yielded larger detection thresholds than regular textures [F(1,31) = 32.3, P < O.OOOS].The effect of element randomization with the two types of texture stimuli used cannot be attributed to the fact that the animals were not trained in the randomized texture conditions, since additional training after the completion of the present experiments did not change the thresholds obtained in these conditions. In summary, for a range of intermediate element to spacings, size thresholds were near-proportional element spacing. The range of element spacings corresponding to this proportional relationship was shifted towards smaller values in the dot/annulus texture stimuli (2.00-0.67 deg) compared to the line texture stimuli (2.67-l .60 deg). At extreme element spacings, this proportional relationship broke down, as indicated by increased relative size thresholds. In addition, detection sensitivity decreased after randomization of the position of the texture elements. Finally, the cats easily generalized to conditions for which they had not been explicitly trained (element spacings 2.67, 1.60,

al.

1.33 and 0.67, as well as the randomized conditions).

Relative length thresholds determined without &urel background reversal Before participating in the experiments described in the previous section, all five cats had participated in a preliminary experiment in which no figure/background reversal was applied to the stimulus presentations. In this preliminary experiment, the stimuli were slides, projected onto a translucent screen. A transition was made from a luminance-defined to a texture-defined target using a fading procedure. In all trials, the target was filled with horizontal line elements and the background was made of vertical ones. Detection performance was measured at element spacings of 4.00,2.00 and 1.31 deg, using regular as well as randomized textures. Interestingly, the optimal performance at intermediate element spacings described in the previous section was also present when figure and background textures were not randomly exchanged [Fig. 4(A)]. In addition, the detection task was solved less efficiently in the randomized texture condition. The summary figure presented in Fig. 3(A) and obtained with random ~gure/background reversal, has been replotted in Fig. 4(B) for easy comparison. Except for the smaller length thresholds in Fig. 4(A) compared to Fig. 4(B) [probably due to the use of horizontal vs vertical line elements to define the target in the experiment illustrated by Fig. 4(A)], the data are remarkably similar. The resemblance between the two data sets should not be taken as evidence that the underlying processes generating the data were identical. Indeed, after the testing, the animals were confronted

CAT 57,59 ,S3

0.3L

1

L 4

1 2 Uemmt

0.3L

8 1 spacing

texture

(viswi

t 4

I

1

2

1

dogroe&

FIGURE 4. (A) Averagerelative length thresholds plotted as a function of element spacing for cats 57, 59 and 63. These thresholds were measured without random reversal of figure and background textures. Line elements in the target were horizontal, those in the background were vertical. The thresholds were measured using the method of constant.&imuli. Proportions of correct responses were measured for a total of 240 trials at relative lengths of 0.38, 0.50, 0.63, and 0.75. After linear regression through the Z-transformed proportions of correct responses, 73.5% correct thresholds were determined in each cat separately, and were then averaged over the cats. Open symbols and stippled lines refer to the randomized texture condition, solid symbols and full lines refer to the regular texture conditions. Error bars represent standard errors. (B) The summary graph of Fig. 3(A) replotted for easy comparison. The data in this figure were obtained with a random trial-to-trial exchange of the textures defining figure and background. The textures used consisted either of left oblique line elements or of right oblique line elements.

TEXTURE

SEGREGATION

with stimuli in which the target was filled with vertical lines and the background with horizontal elements, i.e. the reverse of the stimulus arrangement the animals were used to. In 4 of 5 animals, performance immediately fell to chance. In one animal (cat 65), performance initially remained unaffected after figure/background reversal (De Weerd, Vandenbussche & Orban, 1989), but during further testing, performance dropped to chance level as in the other animals. Hence, there was no transfer of detection performance to a texture stimulus after figure/background reversal, suggesting that the animals in the present experiment identified a particular texture arrangement, rather than texture differences. A

313

IN THE CAT

Size of texture elements at detection threshold as a function of the number of elements in the target square A striking feature of the curves shown in Fig. 3 is the optim~ation of performance at intermediate element spacings. We wondered whether that particular shape of the curve could be attributed entirely to changes in element spacing. Indeed, our arbitrary decision to hold the target area constant with decreasing element spacing implies that an increasing number of texture elements are positioned in the target area during this operation. In the present experiment, we held the number of texture elements in the target square constant while decreasing element spacing, which in this case caused target area to

El

CAT 57,59,63

1.4

CAT 64.65

7.4

target D--Q e--o

16

1.2

A._....*

64

0

256

r

elements 4

1.0

0.8

ov

f

e---O

6.......

06

a

ck2L 1 4

1

L

2

1 spacing

etemeot

ELEMENT 4

SPA{

sual

degrees)

\IG h4sual

degrees) 2

-

16

64

FIGURE 5. Average relative length thresholds (A, cats 57, 59, and 63) and average relative diameter thresholds (B; cats 64 and 65) plotted as a function of element spacing, for different numbers of target elements as indicated in the inset. Only randomized textures have been used in the present experiments. The vertical arrow in (B) indicates that at a 0.50-deg element spacing no 73.5% correct thresholds could be obtained when the number of target elements was reduced to 64. (C) shows line texture stimuli at two of the element spacings and to illustrate the mixed manipulation of both element spacing and number of target elements. Solid symbols in (A) and (B) indicate thresholds obtained with the standard 8 by 8deg target. Standard errors were in the order of 15% of the average threshold.

314

PETER DE WEERD et al.

decrease [Fig. 5(C)]. This experiment was carried out for different numbers of target elements (1,4, 16,64 and 256 elements), for each of the two types of texture stimuli used. Only randomized textures were used. Figure 5(A) shows the average results obtained using the line texture stimuli whereas the average results from the dot/annulus texture stimulus are plotted in Fig. 5(B). These figures show that in addition to element spacing, the number of target elements is an impo~ant determinant of detection performance. Using only the l.OO- and 2.00-deg element spacings for ANOVA, the reduction of the number of target elements significantly decreased detection performance with line texture stimuli [F(2,46) = 22.24, P < 0.0005] and with dot/annulus texture stimuli [ic(2,23) = 30.60, P < O.OOOS].Hence, previous results [Fig. 3(A, B)] obtained with an 8.00 by 8.OOdeg target were determined by both element spacing and the number of target elements. This is well illustrated by the solid symbols in Fig. 5(A) and (B), which correspond to thresholds obtained with the standard target dimensions. Indeed, the superior performance at the 2.00-deg element spacing compared to the 4.00-deg element spacing obtained with line texture stimuli at the standard target area [Fig. 5(A)] must be att~buted to the increased number of target elements, since element spacing did not affect performance for constant numbers of target elements. On the other hand, the increase of the thresholds at the I.OO-deg element spacing compared to the 2.00-deg element spacing was largely due to the element spacing itself, since thresholds increased for all numbers of target elements. The results obtained with the dot/annulus texture stimulus show a similar picture. At the standard target area [solid symbols in Fig. 5(B)], element spacings decreasing till 1 deg were accompanied by decreasing thresholds, but this decrease was mostly due to the increase of the number of target elements. On the other hand, the increase of the thresholds at the smallest element spacing was a genuine element-spacing effect. Notice that the solid symbols in Fig. 5(A) and (B) nicely replicate the results represented in Fig. 3(A) and (B). Although the increased thresholds at the smallest element spacing in Fig. 5(A) and (B) were primarily an effect of element spacing itself, the increase of the thresholds was larger when the target consisted of fewer elements. Stated otherwise, the number of target elements was more important when the elements were small than when they were large. This interaction was si~ifi~nt for the line texture stimulus [F(2,46) = 3.58, P c 0.051. A similar interaction effect took place when using dot/ annulus texture stimuli, since at the 0.50-deg element spacing, we were unable to measure 73.5% correct thresholds if the target consisted of less than 256 elements [indicated by the arrow in Fig. 5(B)]. Finally, for both stimulus types used, targets consisting of only one texture element caused all cats to respond randomly, even at the largest element spacing. In summary, the shape of the curves in Fig. 3 obtained with the standard target area is no simple effect of element spacing. Rather, it reflects an interplay between element spacing and the number of target elements. At

CAT 57,59,63

A

B

regular

randomiSod

1.28

tine width

I;

ah _____ 0.2 ._..__*.. i-J,

1.0

06_

0.31

‘

2

i

,

0.3

J

0.5

FIGURE 6. Average (cats 57, 59,

4

and

63)

‘

2

1

1

1

0.5

relative length thresholds

plotted as a function of eIement spacing, for three different linewidths, as indicated in the inset. The results for the regular texture condition are shown in (A); the results for the randomized texture condition are shown in (B). Error bars represent standard errors. The results shown for the 0.20-deg tinewidth are copied from the average graph in Fig. 3(A).

large and intermediate element spacings, performance with the standard target area was mostly determined by the number of target elements, whereas the effect of element spacing itself became dominant at the smallest element spacings. The influence of Einewidth upon length thresholds Previous experiments have shown that detection thresholds can be obtained at smaller element spacings when using dot/annulus texture stimuli than with line texture stimuli. One explanation could be that decreasing the length of line elements-which are already small at small element spacings--causes a loss of orientation information due to the decrease of the length~width ratio of these line elements. If this were the case, relative length thresholds should decrease when using line elements of O.lO-deg width, especially at small element spacings. On the other hand, when using a 0.40-deg linewidth, thresholds would necessarily rise. Figure 6 shows the average detection performance of cats 57, 59 and 63 for the different linewidths and element spacings, with both regular [Fig. 6(A)] and randomized texture stimuli [Fig. 6(B)]. The average data of Fig. 3(A) obtained at linewidth 0.20 deg (summary figure) are included in Fig. 6 and also in data analysis. The dominant feature of the data is the decreasing detection thresholds with decreasing linewidth [F(2,142) = 12.05, P < 0.0005]. Notice also that the optimum performance at the 2.00-deg element spacing observed for the 0.20-deg linewidth is confirmed at the two other linewidths. At extreme element spacings, thresholds were significantly elevated compared to medium spacings [F(2,142) = 80.06, P < O.OOOS],As in previous experiments, randomizing the position of the texture elements

TEXTURE

SEGREGATION

proved to decrease detection performance [F( 1,142) = 44.07, P c O.OOOS].Figure 6 also suggests that the effect of linewidth was stronger for randomized textures compared to regular textures. However, this interaction was not statistically significant [F(2,142) = 1.961. The preceding data analysis was confined to the 4&O-, 2.00- and l.OO-deg element spacings. For these element spacings, there was a small but consistent effect of linewidth upon detection performance, equal in magnitude at all three element spacings. However, this effect was much larger with the OSO-deg element spacing, since reducing width to 0.10 deg allowed us to measure relative length thresholds, whereas at an identical element spacing no thresholds could be determined when linewidth was 0.20 deg (Fig. 6). This indicates a substantial influence of the length/width ratio upon detection performance at small element spacings. Interestingly, at the 0.50-deg element spacing the effect of randomizing the position of the texture elements was significantIy larger than at the 1.OO-degelement spacing [F&30) = 8.34, P < O.Ol]. To s~ma~ze, length thresholds became smaller with the reduction of linewidth. This effect was largest at the OSO-deg element spacing for both regular and randomized texture stimuli, since no threshold could be determined for linewidths exceeding 0.10 deg. Comparison of relative length thresholds, orientation d@erence threshold and orientation randomization threshold In the present experiment, two alternative detection thresholds were introduced in addition to relative length thresholds. Firstly, orientation differences between line elements of target and background matching the 73.5% correct detection performance were assessed as a function of element spacing (orientation difference thresholds). Secondly, the o~entation randomization ranges corresponding to a 73.5% correct detection performance were determined as a function of element

315

IN THE CAT

spacing (orientation randomization thresholds). In the latter type of thresholds an increase of thresholds indicates better performance. No additional training was required to transfer the animals to these tasks. Figure 7 presents a comparison of average segregation performance measured with orientation difference and o~entation randomization thresholds, with segregation as measured with relative length thresholds. The relative length data [Fig. 8(A)] are the average data presented in Fig. 3(A). At the optimal element spacing of 2.00deg, the average relative length thresholds were O-44 and 0.54 in the regular and randomized texture conditions respectively. At the same spacing, average o~entation difference thresholds were measured of 32.5 and 38.0 deg and orientation randomization thresholds of 20.3 and 38.7 deg, again in regular and randomized texture conditions respectively. Orientation difference thresholds, as relative length thresholds, were significantly affected by changes in element spacing [F(2,46) = 45.74, P < O.OOOS].At the smallest element spacing tested (1 .OOdeg), thresholds increased dramatically compared to those obtained at the 2.00-deg spacing, becoming as large as 50-60 deg [Fig. B(B)]. As observed for relative length thresholds [Fig. 8(A)], orientation difference thresholds determined using the randomized texture were slightly but systematically elevated compared to those obtained using the regular texture [F(1,46) = 15.39, P < O.OOOS].Unlike relative length thresholds, orientation thresholds did not increase at the 4.00-deg element spacing compared to the 2.00-deg element spacing. Detection performance measured with orientation randomization thresholds [Fig. 8(C)] again, depended upon element spacing [E(2,46) = 10.66, P < O.OOOS]. Indeed, detection performance was slightly better at the 2.00-deg element spacing compared to the other two element spacings. In accordance with all previous experiments, detection performance was better at

CAT 57,59,63 A

-

regular

o---a

random.

B

c

loo,

80..

0.6 _

’ & $ 0

601

40_

.-

I

4

2

spacing

(visual

J 1

zi

t

4

1

2

1

degrees)

FIGURE 7. Different assessments of segregation sensitivity as a function of element spacing obtained with line texture stimuli. Data were averaged over cats 57, 59, and 63. (A) Average relative length thresholds plotted as a function of element spacing. The data are copied from the average graph of Fig. 3(A). (B) Average orientation difference thresholds plotted as a function of element spacing. (C) Illustration of the dependence of orientation randomization thresholds upon element spacing. Open symbols and stippled lines refer to the randomized texture condition, solid symbols and fir11lines refer to the regular texture condition. Error bars represent standard errors.

316

PETER DE WEERD ei uf.

regular compared to randomized texture patterns [E;(1,46) = 28.21, P < O.OOOS]. Hence, at smail element spacings, the three measures of detection performance used, revealed decreased segregation performance compared to medium element spacings. However segregation performance measured with orientation difference thresholds was equal at the 4.00- and 2.00-deg element spacings, which suggests that the effect of large spacings upon segregation performance varied with the type of threshold measurement (see Discussion). The importance of texture elements at the border of the target

Figure 5 shows that relative length thresholds depend upon element spacing as well as upon the number of elements within the target. In the present experiment, we wondered whether this effect of the number of elements in the target can be reduced to an effect of target area. In that case, all texture elements in the target would contribute equally to the detection of the target square, and presenting the animals with the target border alone would result in decreased segregation performance. However, the animal presumably detects texture differences, due to the random figure/background reversals, and one can therefore hypothesize that texture elements at the border of the target are most useful for the detection. If only the texture difference at the border of target and background were used, reducing the target to only its outer border, by assigning the background orientation to

all inner target elements, should not diminish performance. We investigated the effect of presenting the target border alone (open square at the abscissa in Fig. 8) by assessing relative length thresholds and o~entation difference thresholds as well as orientation randomization thresholds. Figure 8 compares the average results of this experiment to those results with the complete target obtained previously (Fig. 7). Relative length thresholds [Fig. 8(A, D)], orientation difference thresholds [Fig. 8(B, E)] and o~entation randomi~tion thresholds [Fig. 8(C, F)] are plotted as a function of target type. The three upper graphs [Fig. 8(A-C)] are data obtained at a 2.00-degree element spacing, whereas the graphs below [Fig. 8(D-F)] correspond to data gathered at a l.OO-deg element spacing. Six ANOVAs were carried out, one analysis for each combination of element spacing with type of threshold measurement. The only significant effect to come out of this analysis was the interaction between target type and element position randomization [F( 1,30) = 10.88, P < O.OOS]when relative length thresholds were assessed at an element spacing of 1.OOdeg [Fig. 8(D)]. This effect was due to the sign&ant effect of target type upon thresholds determined with randomized textures [a posteriori T(16,32) = 11.27, P < 0.011. This also entailed the significant increase of the thresholds obtained with randomized compared to regular textures when the target was reduced to its border [a posteriori T(16,32) = 22.54, P < 0.01). It is interesting to compare the data obtained in the randomized texture condition in Fig. 8(D) with results

CAT 57,59,63

B

c

L

P------p I&--% FIGURE 8. Different assessments of segregation sensitivity in the standard target condition (hatched square at the abscissa: dam taken from Fig. 7) and in a condition in which the target was reduced to its outer border (open square at the abscissa). Data were averaged over cats 57,59, and 63. (A, D) Average relative length thresholds plotted as a function of eIement spacing. (B, E) Average orientation difference thresholds presented as a function of element spacing. (C, F) Illustration of the dependence of orientation randomization thresholds upon element spacing. The upper row of plots (A, B, C) corresponds to data obtained at a 2.00-deg element spacing, whereas the lower row of graphs (D, E, F) refers to the l.OO-deg element spacing. Gpen symbols and stippled lines refer to the randomized texture condition, solid symbols and full lines refer to the regular texture conditions. Error bars represent standard errors.

TEXTURE

SEGREGATION

presented in Fig. 5(A), which shows the effect of reducing the number of target elements upon relative length thresholds which were also measured with randomized textures. Reducing the standard target to its border at a l.OO-degree spacing corresponds to a decrease from 64 to 28 elements, and presenting this 28-element border caused a significant increase in relative length thresholds [Fig. 8(D)]. Data presented in Fig. 5(A), on the other hand, show that a much larger decrease in the number of target elements from 64 to 16 did not significantly increase relative length thresholds when using randomized textures [a pusteriori T(24,48) = 2.821. However, in the experiment leading to the results presented in Fig. 5(A), the remaining target elements were grouped together, whereas in the present experiment [Fig. 8(D)] the remaining target elements were dispersed over the border of the target. This suggests that when line elements are small and carry reduced orientation information, a sufficient number of target elements must be grouped together in order to permit the detection of a discontinuity. This conclusion clearly does not apply to regular texture conditions [Fig. 8(D)]. This might be explained by the use of additional cues in the regular texture condition, such as low spatial frequency cues in target and background resulting from the alignment of line elements (see Discussion). In short, replacing the inner target elements with elements from the background did not affect detection performance, whichever threshold type was used to assess detection sensitivity. Hence, the inner target elements contributed nothing to the visibility of the texture difference which the animal was required to detect, with one exception: at a l.OO-deg element spacing, reducing the target to its border caused a significant increase in relative length thresholds when texture element positions were randomized. This suggests that differences between textures composed of small texture elements randomized in position are detected only when a sufficient number of similar elements are grouped at each side of the texture border. DISCUSSION Our data show that cats are able to use texture differences as a cue in a segregation task. That cats can use local discontinuities as a cue in segregation tasks has been demonstrated earlier in the domain of stereoperception (Fox & Blake, 1971; Lepore, Ptito & Lassende, 1986). In the present study, we assessed texture segregation in the cat using two types of texture stimuli and determined the effects of parameter manipulations on segregation performance. We will now summarize the main findings of our study and indicate parallels with human performance as studied by Nothdurft (1985a,b, 199Oa) and Landy and Bergen (1991). Before indicating these lines of parallel evidence, it must be noted that in the justmentioned human studies, the observer’s task was to respond to the shape of a texture-defined region, whereas in the present study the animals did not have to identify the shape of the target. In addition, figure and back-

IN THE CAT

317

ground textures remained identical during testing in the human studies. The efficiency of segregation as measured with size thresholds was a U-shaped function of element spacing in both the line texture and the dot/annulus texture stimuli. At intermediate ranges of element spacing however, length or size thresholds and element spacing were near-proportional to each other (provided target area is kept constant). This relationship between size and element spacing is in agreement with Nothdurft’s (1985b) findings. It suggests that the segregation of two textures remains unaffected for a large range of viewing distances, an inference confirmed in the Landy and Bergen (1991) study. Furthermore, o~entation difference thresholds determined during segregation of line textures were much more elevated than could be expected on the basis of orientation discrimination thresholds determined with single line elements equal to those used in the texture stimulus. Orientation difference thresholds in the segregation task were determined with linelengths equal to 95% of the element spacing. At spacings of 4.00, 2.00 and 1.OOdeg, orientation difference thresholds were 3 1.6, 31.5 and 56.9 deg respectively (averaged over randomized and regular texture conditions). When single line elements were used with lengths of 4.0, 2.0, and 1.0 deg, orientation difference thresholds smaller by a factor of 5 were obtained, being 4.5,6.7 and 12.7 deg respectively (De Weerd et al., 199Ob). Similarly, in humans, typical o~entation thresholds dete~ined with single, short lines (1.0 or 2.0 deg) are in the order of 2 deg (Orban et al., 1984), whereas a IO-deg orientation difference seems to be the under limit for successful segregation performance, whatever the parameters are of the textures used (Nothdurft, 1985b). This finding is in keeping with Nothdurft’s (1985a) observation that the orientation differences between neighboring texture elements can be perceived before the texture border composed of these orientation differences is detectable. Furthermore, a general effect observed in both human and cat experiments is the slight but consistent decrease of segregation performance in all randomized texture conditions (see below). Finally, an interesting finding is that texture elements near the border contribute most to segregation performance, thereby confirming the notion that segregation is based on local di~ontin~ties. The impor~n~ of local transitions in structure (structure gradient) between different textures as a cue for segregation was reported by Nothdurft (1985b) in humans and further corroborated by Landy and Bergen (1991). The similar effects of parametric variations on segregation in cats and humans illustrates that cats can be used as an animal model of human segregation performance. Julesz has proposed that texture segregation is to be explained by first order differences between local features (Julesz, 1980, 1981; Julesz & Bergen, 1983). He referred to these local features as textons. Line crossings, line ends and lines (blobs) of a specific orientation, length, width and color have been identified as textons. The use of physiologi~lly plausible filter models to code textures

318

PETER DE WEERD et al.

has confirmed the importance of local coding in segregation tasks. Bergen and Adelson (1988) showed that the output of simple, size-tuned linear filters could be used to segregate a square filled with crosses from a background of Ls. Changes of the texture elements so as to obtain better or poorer segregation were explained by changes in the outputs of these local filters. Other authors have demonstrated that the output of orientation anisotropic filters (such as oriented Gabor filters) can be used for the segmentation of a variety of texture differences (Fogel & Sagi, 1989; Rubenstein & Sagi, 1990; Sagi, 1990; Turner, 1986). However, the use of the output of the filtering stage supposes a second processing stage in which differences in the dist~bution of local activity are encoded (Nothdurft, 1991). Voorhees and Poggio (1988), in a Marr (1982) inspired view, have proposed a statistical comparison of the attributes (e.g. orientation) of primatives (e.g. blobs) extracted by local filtering. In studies from the Sagi group (Fogel & Sagi, 1989; Rubenstein & Sagi, 1990; Sagi, 1990) it is proposed that the initial filtering is followed by second stage filtering by Laplacian or Gaussian operators, the filters being several times larger than the filters used during first stage filtering. Thus, according to the referenced models, texture segregation requires local filtering (first stage), followed by higher order processes (second stage) required to encode texture differences. From Wilkinson’s results, as well as from our own it can be inferred that the cat visual system can operate in either of two modes to solve the segregation task, and that these two modes might be directly related to the two stages of texture processing described in the previous paragraph. Indeed, in cats which were presented for a long time with the same line texture arrangement, segregation performance fell to chance level after a reversal of figure and background textures. Performance was thus entirely dependent upon the particular arrangement of target and background textures. This may suggest that the animals ignored the background texture, and that detection performance was based solely on the output of one class of local filters signalling a particular orientation. Hence, the output of the initial filtering stage may be sufficient when the task is solved with an identification strategy. On the other hand, when figureand background textures were randomly exchanged, the cats were forced to use the texture difference (differencing strategy). In these circumstances, the stage of initial filtering must be followed by a second stage of difference encoding. The distinction between the two processing modes described is supported by our finding that cats using a differencing strategy continued to segregate textures efficiently even when unfamiliar textures were administered, in contrast to cats using an identification strategy (see Methods and Results). Interestingly, we found that the effects of element spacing and element position randomization on the segregation of line textures were very similar, independent of the strategy used to solve this task. This suggests that effects of global manipulations such as changes in spacing and element position randomization are determined

at the level of initial filtering. Hence, global aspects of the textures seem to be available in the output of the network of initial filters. In general, segregation was optimal at intermediate spacings, and deteriorated at both larger and smaller spacings (see Fig. 7). The decrease of segregation performance at small spacings can be explained by acuity constraints, texture elements becoming too small for the filters available. At larger element spacings, texture elements are larger and are undoubtedly resolved, but their number (and density) may be too small to trigger a segregation process efficiently (Sagi, 1990). The exception-with equally efficient segregation at the 4.00and 2.00-deg spacings-occurred when measuring the segregation of line textures with orientation difference thresholds. Contrary to instances in which segregation was measured with length or orientation randomization thresholds, the assessment of orientation difference thresholds entailed a significant increase of the probability of near-junctions of line elements [see Appendix (6)]. The occurrence of (near)-overlaps between elements might have been especially prominent in conditions of large spacing and large elements, and could have constituted an additional cue used during segregation, thereby explaining why orientation difference thresholds remained constant when element spacing was increased from 2.00 to 4.00 deg. The more efficient segregation performance in regular textures could be related to a contribution of aggregated input to low spatial frequency mechanisms. Indeed, in regular patterns, the texture elements are aligned in rows and columns, which gives the appearance of two gratings of different orientation, and large filters coding the orientation of these “gratings” might give input to the segregation process. The question remains why segregation of dot/annulus texture stimuli could be performed at much smaller spacings than the segregation of line texture stimuli. This difference remained even when the length/width ratio of the line elements was increased. It is possible that a different type of initial filtering is involved in the coding of the dot/annulus texture stimulus. In the first place, it cannot be excluded that simple center-surround sizetuned mechanisms directly contribute to segregation performance (Bergen & Adelson, 1988; Nothdurft, 1990b). Furthermore, the visual system might code differences in the curvature of the circumference of dots and annuli (Versavel, Orham & Lagae, 1990). However, expressing curvature as the inverse of the radius and converting size thresholds into curvature differences, a reduction of element spacing from 4.00 to 0.50 deg was accompanied by an increase in the curvature difference between targetand background elements from 0.64 to 2.47. This shows that the curvature cue is not the major determinant of segregation at threshold level. An additional parameter which could be used as a cue in segregation tasks, is the difference in fundamental spatial frequency content of the dot texture and the annulus texture (Mayhew & Frisby, 1978). In the repetitive pattern of a dot texture, there are as many black/white cycles as there are dots,

TEXTURE SEGREGATION IN THE CAT

whereas this number is twice as large in an annulus texture. Differences in spatial frequency could be encoded locally by Gabor filters, which by definition are also selective for spatial frequency. However, it must be pointed out that for a given element spacing, differences in fundamental spatial frequency are maximal at intermediate relative diameters, and smallest at large and small relative diameters. Hence, if ~rfo~an~ were linked exclusively to the different fundamental spatial frequency contents of target and background, performance would be best at inte~ediate relative diameters, and worst at small and large diameters. This was not the case, and therefore, differences in fundamental spatial frequency alone cannot explain the results obtained. One could speculate that segregation performance in the dotlannulus texture stimulus might be mediated by more than one type of filter, and that the combined output of these filters accounts for the accurate segregation of dot/annulus textures at small element spacings. Whatever the type of filter is contributing to segregation of dot/annulus textures, it is important to realize that it is unnecessary to resolve both the dots and the annuli in order to segregate the dot/annulus texture stimulus. For instance, even if the spatial frequency content of the annulus texture surpasses the acuity limit, the dot texture will be resolved, and as long as this is the case, accurate detection performance can be expected. Indeed, the annulus texture would then appear homogeneous whereas a vague impression of granularity would remain in regions occupied by the dot texture. The much weaker acuity constraints imposed by dot/annulus textures compared to line textures might be the major factor explaining the robustness of segregation of the former type of texture stimuli at small element spacings. The initial filtering of textures is essentially local in nature. Our finding that segregation performance is mediated mainly by texture differences at the border between target and background, as well as related findings in humans (Landy & Bergen, 1991; Nothdurft, 1985b), suggests that the operations required to encode texture differences are also local in nature and not necessarily as global as has been suggested by the Sagi group (Fogel & Sagi, 1989; Rubenstein & Sagi, 1990; Sagi, 1990). Especially in the conditions in which texture elements within figure and background areas were equal, the task, in principle, could be solved by merely detecting a difference between two neighboring elements. In other conditions however, this strategy cannot work. For example, when randomizing the orientation of line elements, keeping the average orientation difference between figure and ground at 90 deg [see Appendix (3)], more than two elements have to be compared in order to reliably distinguish different textures. Indeed, for the two textures involved, an estimate of the variability in the output between the filters carrying the orientationsignal must be generated before a comparison between figure and background can be carried out. Note that variability of the orientation signal can also be a consequence of element position randomization. Rubenstein and Sagi (1990) have shown that irregularities in the

319

spacing of (parts of) texture elements can cause variability between the outputs of individual Gabor filters encoding the textures. Finally, making the size of the elements small causes noisy output of individual filters. In all these cases, one would expect that a larger sample of elements of target and background must be compared in order to segregate the textures, and therefore, under these conditions, reducing the target to its border should reduce the efficiency of texture segregation. This does not happen (except in a combination of short line elements and small spacing with the randomized texture condition), which at first sight might seem somewhat paradoxical, since efficient segregation based on variable filter output must require a compa~son of the output of a sufhcient number of filters (subserving different loci). The paradox can be resolved if one assumes that in the second stage of texture processing, the visual system directly encodes borders, rather than encoding isolated texture differences. Our finding that cats perform randomly if the target consists of only a single element, even when the texture elements are large, supports this hypothesis. Conversely, the less efficient segregation in conditions of element orientation randomization and element position randomization suggests that the visual system pays the price of less efficient segregation rather than increasing the region from which texture boundaries are reconstructed. Nothdurft’s (1985a) finding that small orientation differences at the border between two textures or line elements can be observed without actually detecting the border, as well as related findings (see De Bruyn & Orban, 1988; van Doorn & Koenderink, 1983) suggest that only sufficiently large differences give rise to boundary designation. Orban and Gulyhs (1988) explained this by proposing that difference detectors (or boundary detectors) receive only coarse filter output and can therefore signal only coarse differences. It is behaviorally plausible that the visual system is so designed, since differences are most likely to be large at the border between two distinct surfaces. The primary importance of the difference at texture borders reported in the present study, by Nothdurft (1985b), and Landy and Bergen (1991) fits with such a system design. On the other hand, the fact that the finest estimates of orientation difference are not used for boundary construction suggests that the coding of fine differences is dealt with by a separate process, called the continuity stream by Orban and Gulybs (1988). In the continuity stream, the coarse output of a number of filters (local detectors) is combined in order to compute very fine estimates of visual parameter values (Orban, Vandenbussche, Sprague & De Weerd, 1990; Vogels, 1990). Contrary to the discontinuity stream which is involved in the detection and processing of bondaries between surfaces, the encoding of visual parameters in the continuity stream would contribute to the detection of the local structure of surfaces within the boundaries delineated by the discontinuity stream. A classical distinction, having its origins in work with visual search tasks and texture segmentation tasks in humans, is that between parallel (preattentive) and serial

320

PETER DE WEERD et al.

(attentive) processing (Julesz, 1984; Julesz & Bergen, 1983; Treisman & Gelade, 1980). Parallel processing allows the detection of a texture difference, but the texture elements which give rise to this difference are not perceived at that stage (Julesz, 1986). Parallel processing directs attention to the locus of the difference, and serial inspection allows the identifi~tion of texture elements and the construction of boundaries. It is difficult to relate the attentive/pre-attentive distinction (Julesz, 1986) to the distinction between the continuity and discontinuity streams proposed by Orban and Guly& (1988). Neither do our data give any clue as to which of these processes are invoked under the different conditions of our experiments. Fu~he~ore, although the present experiments suggest that cats using a differencing strategy for segregation may use texture borders rather than isolated texture differences, this evidence remains indirect. Our experiments also give no answer to the question of whether cats are also able to process attributes (e.g. orientation) of texture borders. A largely unaddressed issue concerns the neural substrate of texture segregation. The receptive fields involved in initial filtering of textures should be small, and it is therefore unlikely that cells outside areas 17, 18 and 19 contribute to the local coding of texture elements (Palmer, Rosenquist & Tusa, 1978; Tusa & Palmer, 1980). In contrast, the selectivity of many cells in area 17, 18 and 19 for orientation, length, width and spatial frequency as well as their small receptive fields indicates that they could be involved in the local coding of texture elements (Albrecht & De Valois, 1981; Kato, Bishop & Orban, 1978; Movshon, Thompson & Tolhursh, 1978; Orban, 1984; Rose, 1977; Saito et al., 1988). As far as Gabor filtering is involved in texture segregation, it is likely that S-cells in area 17 and possibly area 18 are the substrate of that operation (Jones & Palmer, 1987; Palmer, Jones & Stepnoski, 1991). Furthermore, observations by Nothdurft (1990b) on the processing of textures by units in the cat’s Lateral Geniculate Nucleus (LGN) have shown that these units are useful in the initial filtering of a variety of textures, possibly including our dot/annulus texture stimuli, but probably not line texture stimuli. A limited number of observations are available about the manner in which the visual system may represent texture differences. In the motion domain, Orban, Guly& and Vogels (1987) and GulyBs, Spileers and Orban (1990) have recorded cells in cat areas 17 and 18 which signalled opposite direction of motion between a light bar moving in a particular direction and a moving noise background moving in the opposite direction (antiphase conditional cells). A number of cells did so inde~ndently of whether the light bar moved in the preferred or in the nonpreferred direction (antiphase cells). The former cells signal signed differences in the direction of motion, the latter ones signal unsigned dzxerences. Detectors of the latter type are assumed to be involved when cats use a differencing strategy. Interactions between moving bars and textured backgrounds similar to those reported by Orban et al. (1987) and

Gulyas et al. (1990) have been reported by von Grunau and Frost (1983) in visual Lateral Suprasylvian areas of the cat. Unsigned difference detectors more relevant to the textures used in the present study were discovered by Knierim and Van Essen (1989). These authors have reported units in monkey VI in which the response to a line segment centered in the receptive field was enhanced by a surround containing line segments at an orthogonal orientation, irrespective of the central orientation. Such detectors have not yet been reported in the cat (see Nothdurft & Li, 1984). The location of neural substrates contributing to texture segregation can only be identified through lesion ex~riments. In the monkey, Schiller, Logothetis and Charles (1990) have shown that the segregation of a line texture stimulus similar to ours is heavily impaired after parvocellular lesions in the LGN. Since spatial tasks requiring high resolution were systematically affected following this lesion, this suggests that adequate resolution is a key precondition for accurate texture perception. Since the X-system in the cat is best suited for the transmission of high resolution info~ation (Lennie, 1980; Orban, 1984), and since area 17 is the sole target of the X-channel (Sherman, 1985), it is possible that area 17 of the cat-with its high incidence of X-fed Gaborlike units (Jones & Palmer, 1987+is of major importance for the initial filtering of fine textures. Since the selectivity of area 18 units is limited to the lower spatial frequency range (Movshon et al., 1978), the cont~bution of area 18 to initial filtering of textures could be restricted to instances of larger elements and larger element spacings. These hypotheses, as well as our intuition that texture differences or texture borders are encoded in areas outside areas 17 and 18, are being tested in current lesion experiments. REFERENCES Albrecht, D. G. & De Valois, R. L. (1981). Striate cortex responses to periodic patterns with and without the fundamental harmonics. Journal of Physiology,

London, 319, 491-514.

Beck, J. (1982). Textural segmentation. In Beck, J. (Ed.), Organization and representation in perception (pp. 285-317). Hillsdale, N.Y.: Erlbaum. Beck, J. (1983). Textural segmentation, second-order statistics, and textural elements. Biotogicul Cybernezics, 48, 125-130. Bergen, J. R. & Adelson, E. H. (1988). Early vision and texture perception. Nature, London, 333, 363-364. Bergen, J. R. & Julesz, B. (1983). Parallel versus serial processing in rapid pattern discrimination Nature, London, 303, 696-698. Berkley, M. A. (1970). Visual discriminations in the cat. In Stebbins, W. (Ed.), Animalpsychophysics (pp. 231--247). New York: Appleton Century Crofts. Caelli, T. & Julesz, B. (1978). On perceptual analyzers. underlying visual texture discrimination. Part I. ~iotogi~ai Cybemefics, 28, 167-175. Caelli, T., Julesz, B. & Gilbert, E. (1978). On perceptual analyzers underlying visual texture discrimination. Part II. Biologicaf Cybernetics,

29, 201-214.

De Bruyn, B. & Orban, G. A. (1988). Human velocity and direction discrimination measured with random dot patterns. Vision Research, 28, 1323-1335. De Weerd, P., Vandenbussche, E. & Orban, G. A. (1989). Texture segmentation in the cat. Archives ~nfernariunafe.~ de Physioiogie et de Biochj~ie,

97, 33.

TEXTURE

SEGREGATION

De Weerd, P., Vandenbussche, E. & Orban, G. A. (1990a). Speeding up visual discrimination learning in cats by differential exposure of positive and negative stimuli. Behavioral Brain Research, 36, 1-12. De Weerd, P., Vandenbussche, E. & Orban, G. A. (1990b). Bar orientation discrimination in the cat. Visual Neuroscience, 4, 257-268. Fogel, I. & Sagi, D. (1989). Gabor filters as texture discriminator. Biological Cybernetics, 61, 103-l 13. Fox, R. & Blake, R. R. (1971). Stereoscopic vision in the cat. Nature, London, 233, 55-56. Gulyas, B., Spileers, W. & Orban, G. A. (1990). Modulation by a moving texture of cat area 18 neuron responses to moving bars. Journal of Neurophysiology, 63, 404-423. Jones, J. P. & Palmer, L. A. (1987). The two-dimensional structure of simple receptive fields in cat striate cortex. Journal of Neurophysiology, 58, 1187-1211. Julesz, B. (1980). Spatial nonhnearities in the instantaneous perception of textures with identical power spectra. Philosophical Transactions of the Royal Society of London, B, 290, 83-94. Julesz, B. (1981). A theory of preattentive texture discrimination based on first-order statistics of textons. Biological Cybernetics, 41, 131-138. Julesz, B. (1984). A brief outline of the texton theory of human vision. Trends in Neuroscience, 7, 41-45. Julesz, B. (1986). Texton gradients: The texton theory revisited. Biological Cybernetics, 54, 245-25 1. Julesz, B. & Bergen, J. R. (1983). Textons, the fundamental elements in preattentive vision and the perception of textures. The Bell System Technical Journal, 62, 1619-1645. Kato, H., Bishop, P. 0. & Orban, G. A. (1978). Hypercomplex and simple/complex classification in cat striate cortex. Journal of Neurophysiology, 41, 1071-1095. Kirk, R. (1968). Experimenial design procedures for the behavioral sciences. Belmont: Brooks Cole. Knierim, J. J. & Van Essen, J. C. (1989). Single unit responses to texture patterns in area Vl of the alert monkey. Society for Neuroscience Abstracts, 15, 323. Landy, S. & Bergen, J. R. (1991). Texture segregation and orientation gradient. Vision Research, 4, 679691. Lennie, P. (1980). Parallel pathways: a review. Vision Research, 20, 561-594. Lcpore, F., Ptito, M. & Lassonde, M. (1986). Stereoperception in cats following section of the corpus callosum and/or the optic chiasma. Experimental Brain Research, 61, 258-264. Marr, D. (1982). Vision: A compulational investigation into the human represenfation and processing of visual information. San Francisco: Freeman. Mayhew, J. E. W. & Frisby, J. P. (1978). Texture analysis and Fourier analysis in human vision. Nature, London, 275, 4388439. Movshon, J. A., Thompson, I. D. & Tolhurst, D. J. (1978). Spatial and temporal contrast sensitivity of neurones in areas 17 and 18 of the cat’s visual cortex. Journal of Physiology, London, 283, 101-120. Nothdurft, H. C. (1985a). Orientation sensitivity and texture segmentation in patterns with different line orientations. Vision Research, 25, 551-560. Nothdurft, H. C. (1985b). Sensitivity for structure gradient in texture discrimination tasks. Vision Research, 25, 195771968. Nothdurft, H. C. (1990a). Texture segmentation is based on local dissimilarities. Investigative Ophthalmology and Visual Science, 31, 561. Nothdurft, H. C. (1990b). Texture discrimination by cells in the cat lateral geniculate nucleus. Experimemal Brain Research, 82, 48-66. Nothdurft, H. C. (1991). Texture segmentation and pop-out from orientation contrast. Vision Research, 31, 1073-1078. Nothdurft, H. C. & Li, C. Y. (1984). Representation of spatial details in textured patterns by cells of the cat striate cortex. Experimental Brain Research, 57, 9921. Orban, G. A. (1984). Neuronal operations in (he visual cortex. Berlin: Springer. Orban, G. A. & Gulyas, B. (1988). Image segregation by motion: Cortical mechanisms and implementation in neural networks. In

IN THE

CAT

321

Eckmiller, R. & v.d. Malsburg, C. (Bds), Neural computers (NATO AS1 Series F) (pp. 149-158). Berlin: Springer. Orban, G. A., Gulyas, B. & Vogels, R. (1987). Influence of a moving textured background on direction selectivity of cat striate neurons. Journal of Neurophysiology, 57, 1792-1812. Orban, G. A., Vandenbussche, E. & Vogels, R. (1984). Human orientation discrimination tested with long stimuli. Vision Research, 24, 121-128. Orban, G. A., Vandenbussche, E., Sprague, J. M. & De Weerd, P. (1990). Orientation discrimination in the cat: A distributed function. Proceedings of the National Academy of Sciences of the U.S.A., 87, 1134-1138. Palmer, L. A., Jones, J. P. & Stepnoski, R. A. (1991). Striate receptive fields as linear filters: Characterisation in two dimensions of space. In Cronly-Dillon, J. (Gen. Ed.) & Leventhal, A. G. (Ed.), Vision and visual dysfunction, Vol. 4, The neural basis of visual function (pp. 246-260). London: Macmillan. Palmer, L. A., Rosenquist, A. C. & Tusa, R. J. (1978). The retinotopic organization of lateral suprasylvian visual areas in the cat. Journal of Comparative Neurology, 177, 237-256. Rose, D. (1977). Responses of single units in cat visual cortex to moving bars of light as a function of bar length. Journal of Physiology, London, 271, l-23. Rubenstein, B. S. & Sagi, D. (1990). Spatial variability as a limiting factor in texture-discrimination tasks: Implications for performance asymmetries. Journal of the Optical Society of America, A, 7, 1632-1643. Sagi, D. (1990). Detection of an orientation singularity in Gabor textures: Effect of signal density and spatial frequency. Vision Research, 30, 1377-1388. Saito, H., Tanaka, K., Fukada, Y. & Oyamada, H. (1988). Analysis of discontinuity in visual contours in area 19 of the cat. Journal of Neuroscience, 8, 1131-l 143. Schiller, P., Logothetis, N. K. & Charles, E. R. (1990). Role of the color-opponent and broad-band channels in vision. Visual Neuroscience, 5, 321-346. Sherman, S. M. (1985). Functional organisation of the W-, X- and Y-cell pathway in the cat: A review and hypothesis. In Sprague, J. M. & Epstein, A. N. (Eds), Progress in psychobiological and physiological psychology (Vol. 11, pp. 233-314). New York: Academic Press. Treisman, A. & Gelade, G. (1980). A feature integration theory of attention. Cognitive Psychology, 12, 97-136. Turner, M. R. (1986). Texture discimination by Gabor functions. Biological Cybernetics, 55, 71-82. Tusa, R. J. & Palmer, L. A. (1980). Retinotopic organization of areas 20 and 21 in the cat. Journal of Comparative Neurology, 193, 147-164. van Doom, A. J. & Koenderink, J. J. (1983). Detectability of velocity gradients in moving random dot patterns. Vision Research, 23, 799-804. Versavel, M., Orban, G. A. & Lagae, L. (1990). Responses of visual cortical neurones to curved stimuli and chevrons. Vision Research, 30, 235-248. Vogels, R. (1990). Population coding of stimulus orientation by striate cortical cells. Biological Cybernetics, 64, 25-3 1. von Grunau, M. & Frost, B. J. (1983). Double-opponent-process mechanism underlying RF-structure of directionally specific cells of cat lateral suprasylvian area. Experimental Brain Research, 49, 84-92. Voorhees, H. & Poggio, T. (1988). Computing texture boundaries from images. Nature, London, 333, 364367. Wetherill, G. B. & Levitt, R. (1965). Sequential estimation of points on a psychometrical function. British Journal of Mathematical and Statistical Psychology, 18, I-10. Wilkinson, F. (1986). Visual texture segmentation in cats. Behavioural Brain Research, 19, 71-82.

Acknowledgements-We kindly acknowledge the critical reading of Steven Raiguel and Rufin Vogels. We are grateful for the expert technical assistance of P. Kayenbergh, G. Meulemans and Y. Celis.

322

PETER

DE WEERD

Finally, C. Fransen was indispensable in executing the experiments. This research was partly supported by grants from the National Research Council from Belgium to G. A. Orban and E. Vandenbussche. P. De Weerd was a Junior Research Fellow of the Nationaal Fonds voor Wetenschappelijk Onderzoek and presently holds a Junior Fellowship at the KU Leuven.

APPENDIX (1) In our experiments, we wished the difference between target and background to be determined by differences other than global luminance differences. Such luminance differences can be associated with small differences in areas between texture elements in figure and background, but they can also be due to artefacts of the display [see Appendix (4)]. (2) When randomizing the positions of line elements, a procedure was used which minimized overlap in the following way: Before a line element was drawn, it was first determined that its randomly chosen position would not interfere with the previously drawn, neighboring elements. If overlap was detected, a new random position was again checked for overlap. This procedure continued until either a nonoverlapping position was found, or to maximum of 100 trials after which overlap was tolerated. When randomizing the positions of dots and annuli, no overlap was allowed. Hence, for annulus diameters larger than 50% of the element spacing, the maximal displacement of the elements was reduced by the difference between half of the element spacing and the annulus diameter. (3) The reduction in the diameter or length (element size) of the texture elements was done by dividing the previous length or diameter by a factor of 1.1. When reducing the orientation difference, a factor of 1.2 was used. The increase of the randomization of the orientation of the line elements was achieved as follows: The smallest orientation randomization range was 2deg. Here, the orientation of each line element was randomly decremented or incremented by 1 deg. The randomization range then was increased in subsequent steps by multiplying it by a factor of 1.25. In this operation, the average orientation difference between line elements in figure and background was held at 90 deg, although orientations near the average itself never appeared, since the actual orientation of each line element was chosen at random from only the two extremes of the randomization range. (4) Our choice of oblique line elements was imposed by an artefact of the screen: horizontal line elements were slightly more luminous than vertical ones, even though their surfaces were equal. This artefact was eliminated by using oblique line elements. For the same reason, reducing the orientation difference between target and background elements (see below) was achieved by manipulating the element orientations in both target and background, choosing these orientations in a mirror-symmetrical way around the horizontal [Fig. l(E)]. This also prevents the animals from using different indentations at the borders of line elements associated with nonsymmetrical orientations. Furthermore, it is worth mentioning that drawing dots and annuli of perfectly equal surface on a digitized image screen poses special

et al.

difficulties, especially when these elements are small. In order to draw annuli and dots of equal surface, the annuli were given an inner diameter equal to the outer diameter of the dots and an outer diameter equal to ,/2 times the inner diameter. However, due to inevitable rounding of floating point values to discrete values, there often is a difference of one pixel between the surfaces of individual dots and annuli. Luminance differences due to a one pixel difference are negligible for individual texture elements, but since in the texture stimulus dots and annuli are grouped in a target and a background area, the cumulative difference can become important. Hence, a calibration procedure was developed in which pixels could be added to or subtracted from the annuh, until the overall luminances of figure and background textures were equal. To reliably measure the luminance of figure and background textures, these textures were generated separately on the screen and their global luminance was measured using a Minolta luminance meter. After this procedure, residual global luminance differences were in the order of 0.01 cd/m2. (5) The DEM training required to make the cats familiar with the trial-to-trial random reversals of figure and background textures has been carried out with horizontal and vertical line elements. After a period of testing with these stimuli, we introduced the use of left oblique versus right oblique line elements to define the texture difference. Although the transfer was successful, and allowed the measurement of segregation performance immediately after the transfer, additional training was given in each condition of element spacing to assess whether the transfer had been complete. Segregation performance as measured with element length thresholds remained unchanged after retraining, which confirms that transfer had been complete, and that the initial global luminance difference between target and background (in the order of 0.1 cd/m2), due to the use of horizontal and vertical line elements [see Appendix (4)], had not interfered with the acquisition of texture segregation. (6) Alternative cues at the border between textures can contribute to segregation performance. Local luminance differences at the border available in regular textures are eliminated by element position randomization. Another cue consists of junctions or near-junctions of texture elements. Measuring segregation performance by means of size thresholds, however, avoids this cue to be significant. Indeed, the magnitude of alternative cues ,decreases with reductions of relative length. Therefore, the measurement of relative length thresholds is the most valid way to assess segregation performance of line textures. In contrast, when reducing the orientation difference between elements of figure and background in line texture stimuli, the incidence of near-junctions of elements at the texture border must increase, because of the mirror symmetrical way in which this manipulation is carried out [see Appendix (4)]. This cue could contribute to segregation performance when measured with orientation difference thresholds. The same holds for cases in which segregation is measured with orientation randomization thresholds, but to a much lesser extent since the orientation randomization avoids line ends of line elements of target and background to approach each other in a systematic way.