Efficient pattern recognition in fine particle science

Efficient pattern recognition in fine particle science

Pattern Recognerion Pergamon Press 1972 . VoL 4 . pp . 147-154 Pnnted us Great Bntam Efficient Pattern Recognition in Fine Particle Science BRIAN H ...

790KB Sizes 0 Downloads 43 Views

Pattern Recognerion Pergamon Press 1972 . VoL 4 . pp . 147-154 Pnnted us Great Bntam

Efficient Pattern Recognition in Fine Particle Science BRIAN H . KAYE Laurentian University, Sudbury, Ontario, Canada (Received 14 December 1970 and in revised form 28 September 19711 Abstract-Many of the problems of Fine Particle Analysis can be tackled using the technology of pattern recognition procedures . In this review the various areas of interest to the Fine Particle analyst of Pattern recognition procedures are outlined and recent research developments are indicated and reviewed .

of papers brought together in this special issue of Pattern Recognition entitled "Pattern Recognition in the Fine Particle Science" were generated as a result of the one day International Seminar held at Laurentian University under the general topic "Computer Interfaces in the Evaluation of Fine Particle Systems" . Fine Particle Science and Technology is a new discipline emerging from the confluence of Physics, Chemistry, Mathematics and Engineering, utilized in the study of materials existing in the finely divided state . The special properties of fine particle systems arise from the fact that the surface to mass forces operative within the system are in competition with each other in a way not noticed at the macroscopic Physics level. In colloid science- surface properties dominate ; in macroscopic physics . mass forces dominate : fine particle science lies between colloid science and traditional physics of materials . An important area of endeavor in fine particle science is the evaluation of the size and shape of particle profiles as viewed through a microscope system or as recorded on a photographic plate. For example, when undertaking the evaluation of the particles present in urban air, the fine particle scientist may find himself in a position of having to evaluate the dust deposited on 1000 microscope slides . In the monitoring of the air in the working area of a mine the scientists may find himself having to look at particles deposited on many membrane filters . The group of Fine Particle Scientists working at Laurentian University were actively engaged in this type of problem and were being swamped by the deluge of data which needed to be evaluated before they could look for patterns of meaning in the deposition rates of dust fallout . Working on the simple premise that the recognition of profiles such as those shown in Fig . 1, represented a relatively simple problem compared to that of recognizing the structure of cursive handwriting, and similar problems studied by scientists working in optical character recognition and in digital evaluation of photographs, we turned to the literature on these topics to see if there was any applicable technology which could be utilized by the Fine Particle Scientists . We were almost overwhelmed by the wealth of data in this area ." 2 ' The level of sophistication employed in the logic operations made it difficult to evaluate the relevance of the technology generated in the area of Pattern Recognition for the Fine Particle Analyst . It was therefore decided to bring together representatives of both communities in a seminar atmosphere to discuss the problem of transferring the technology from Pattern Recognition to the area of Fine Particle Science . The speakers THE SET

!47



148

BRIAN H . KAYE

I 41,

1 i 40

V ** 49 A .4 FIG . 1 . A

94

V

"Fine particle array" .

who took part in this original meeting were, Messrs . Cheng, Medalia, Morton, Shelman and White along with personnel from Laurentian University . All of these people have contributed to this special issue except Dr . Morton and Dr. Chan& Fortunately Dr. Morton's work is available in the commercial literature because of his connection with the rr.M.C. Automatic Microscope System marketed by Bausch and Lomb." It is hoped that this set of papers will bring together the two communities and stimulate mutual research . In a recent review of some Pattern Recognition instruments Dr . Fellgett of Reading University, England, characterized pattern recognition as a subject where almost everything looks promising at first and almost nothing succeeds . Another verdict which has been passed on Pattern Recognition is that it is a set of sophisticated solutions looking for problems . Both of these generalizations have sufficient grains of truth in them to make them worth noting . We hope this series will help to match solutions to the problems of the Fine Particle Analyst. A study of the literature of pattern recognition seems to leave out an important concept which should be part of the working philosophy of Fine Particle Scientists entering this field . It is the concept of "Efficient Pattern Recognition" . We define Efficient Pattern Recognition as sufficient sampling of the available information to achieve the significant decision . This is one of many concepts which is easier to define than to achieve in any practical system . What tends to happen with an array such as that of Fig . I is that the electronic engineer moves in on the system, examines it with a high degree of resolution, transfers the information to a digital computer and utilizes tremendous memory capacity to process all



Efficient pattern recognition in fine particle science

149

the information to achieve a multiplicity of parameters which can be used to describe the system . He tends to go for "overkill" in the logic because he is not too sure of the level of information that any working scientist who will ultimately utilize his information processing machine will require . One of the problems faced by the engineer is that he is not able to recognize the physically significant system from the abstraction represented by an array such as that of Fig. 1 . 1 " Alternatively, because the system of Fig . 1 will .. for example . be given to him as a set of pores in a ceramic body he tends to have a fixation that he is studying ceramic science instead of realizing that the abstracted problem of studying the shape, size, number and location of the profiles in Fig . 1 has implications for many technologists and that if he explores the literature of these allied areas, he may discover useful information . Since the purpose of this set of papers is to bring together different communities to study their mutual problems it is perhaps useful to briefly set out the many types of fine particle systems which with slight variations in shape and density of coverage could be represented by a schematic diagram similar to Fig . 1 . Thus Fig . 1 could represent a sprinkling of pollen grains on a microscope slide, aerosols or mist particles on a membrane filter, air pollution particles on a filter paper, grains of an alloy in a metal section . rock species in an ore section . sections through a ceramic, sections through living tissues .. photographs of chromosomes, stars in the sky, moon craters, trees photographed from an aeroplane or oil slicks on a lake surface . Having sketched out the areas of technology where the pattern recognition expert may look for problems awaiting solutions it is perhaps appropriate to indicate that the fine particle analyst unfamiliar with the literature on pattern recognition will find the useful introduction to the concept of the technology in the article "How we find patterns" by Giuliano and the article "Whatever happened to Cybernetics?"'` 61 In Fig. 2 we show the diagram which was used to publicize the Conference from which these scientific papers have been generated . It is a computerized version of a standard double picture used in psychological testing schemes . The straightforward version of this diagram is known as Rubins Goblet . The picture can be used to represent either a face to face dialogue or a passive vase . Thus the pattern recognized in the picture depends upon

54067 82290 24323 8413

1

I I 1

i 0 63 372 162

1 40 946 64 54 8

1

11 I 7533 6 00607

4 19107

FIG . 2 . Computer card random-number table version of "Rubins Goblet' .



150

BRIAN ti . KAYE

the interest of the viewer . In a very intimate sense both pictures are in the diagram . However, to retrieve one pattern one definitely needs to know the interest of the viewer . This is symbolic of the problems in particle size analysis . The retrieval of the significant information requires a great deal of knowledge about Fine Particle systems . Also in this Fig . 2 we notice that there are many extraneous features which would completely confuse the logic operation and one would need to consider the possibility of pre-processing the picture to avoid the noise level generated by the random digits and all the holes in the dark regions of the picture. As a generalization it would appear to be true that much of the work in pattern recognition has concentrated on taking the information from a picture such as that of Fig . 2 by processing it with some scanning device and then performing a multiplicity of operations on the digital information fed to a computer . It seems that in future research perhaps more effort will have to go into pre-processing of the image to simplify the logic operations and to achieve "Efficient Pattern Recognition" . One of the techniques which appears to be fairly promising as a data compression procedure is spatial filtering of an image of the profiles to be evaluated ."" In Fig . 3 we show the effect of spatial filtering on a diagram when the type of filtering used is either high or low frequency filtration . If one were to be seeking for the contours of a vase of Fig . 2 it could be that the removal of low frequency signals could result in an enhancement of the boundaries of the vase which would be simpler to pick up with any subsequent interrogation logic . If, however, we wish to smooth out the integer background from the random numbers forming the vase we may want to filter the high frequencies to remove the equivalent of the grain of the picture . Alternatively we may be interested in the orientation of the array of small rectangles, in which case some mass optical processing of the field of view may be appropriate . Thus in Fig . 4 an example from a paper by Redman and Reid is shown which demonstrates how optical processing of an entire field of view may yield information on the orientation of particles in a field of view .'"' Another aspect of ahy ultimate implementation of pattern recognition technology in Fine Particle Science which has to be recognized is that sometimes it is extremely difficult to isolate the particles which are of interest . A great deal of expense in the automated microscope systems which have been reported up to this time hinge on the necessity for separating out contiguous particles from the agglomerates .'"' Also the automated procedures currently available are not well adapted to the high level scrutiny of individual profiles. An exception to this statement is the executive light pen feature of the instrument designed by Dr. Morton."' It would appear that future research will concentrate its effort into better procedures for isolating the profile of interest and transferring it to an interrogation system . In Fig . 5 we show a new system under development at Laurentian University which can be used to transfer an image of interest to a subsidiary logic system . The fiber optic bundle of Fig. 5a is taking a profile to a photocell so that the area of the profile may be evaluated by measuring the optical intensity level over the end of the bundle . For this type of measurement, a randomized set of fibers is a sufficient transferal bundle . If, however, one wanted to explore in some detail the structure of an individual profile the system of Fig . 5b could be used. The profile is picked up by a coherent bundle and transferred to a television camera The profile is then displayed on a T .V. monitor, as shown . Another logic scanner could then be used on the image as shown on the screen, with the image defocused to blend the elements of the picture . However, we prefer to use a composite fiber bundle such that half of the fibers go to the screen to enable us to see a

Removal of low frequencies

b Removal of high frequencies

Fie . 3 . Examples of the effect of spatial filtering of an image . Pictures reproduced from Ref . 8 by permission .

[facing page

150)

(b) is the (d) is the

FOURIER FouRIER

transform of the field of view (a) transform of the field of view (c)

FiG . 4 . Examples of the information concerning particle size and particle orientation which can be discerned in the Optical Fourier transform .

Particle area measurement using a photometer and randomized bundle

Coherent bundle used to transfe • image for digital evaluation FIG . 5.

Examples of the use of fibre optic bundles to isolate and transfer a particle image for evaluation . From Ref. 8 .



Efficient pattern recognition

in

fine panicle science

ur

digitized representation of a profile . The other portion of the bundle is splayed out into an interrogation circle and by scanning this circle a digital representation of the profile can be transferred direct to a computer . Thus in Fig . 6 we show a digital representation of a profile which could be stored with very low memory requirements in a digital computer . In fact each element of the fiber optic bundle carries out a local integration of the signal and compresses the data into a finite item of information from a specific area of the image . When the profile has been isolated and compressed into the digital form of the top element of Fig . 6, it is a relatively easy matter to calculate the area of the profile . It is also relatively easy to calculate a maximum cord . It is a simple feature of operator procedure to align a profile as in Fig . 5b with its -maximum diameter along the horizontal direction . The computer can readily search the image for maximum and minimum diameters in this position if the operator carries out this adjustment. Although a more sophisticated logic could be evolved so that the operator did not have to make the decision to line the profile horizontally, we have a minimum effort on the part of the computer if the operator carries out this simple adjustment . Again- if we have stored an image profile in the way shown in Fig . 6 we can readily carry out the calculations of the shape factors which have been suggested by MEDALIA ."' Another technique which we are developing at Laurentian is to automate decisions concerning the probable structure of a profile such as Fig . 6 . We use stripping routines which have been reported in the literature to strip the profile down as illustrated in the subsequent elements of Fig . 6 . We then show that it is reasonable to suspect that in specific cases that a profile is an agglomerate if the stripping routine reduces it to several independent centers as shown in Fig . 6 . One would have to have some general information of the type of particles to be encountered in the profile array before one could utilize this routine to achieve decision making procedures : however. the very low digital representation of the particle as shown in Fig . 6 makes it feasible to consider the stripping routine without excessive demands on the memory core of the computer . Thus efficient pattern recognition will often require pre-processing as a successful element in any ultimate analytical procedure . We are exploring the possibility of storing digital pictures of the type shown in Fig . 6, on a punch tape buffer memory . We then envisage unloading this type of information into a time shared computer. In this way the operations of selection . isolation . transference and digitization which are the relatively slow elements of the procedure can proceed at the operators best speed and then the computer can process all the information in one batch . We envisage that this kind of transferal mechanism can be carried out remote from the data processing . s o that an operator may be able to send his information via punch tape to a time shared computer and receive the print-out of his distribution function . Returning to the concept of efficient pattern recognition, another type of study which deserves attention is the minimum level of digitization which is consistent with the required information . In Fig . 7, we show an example of low resolution sampling from a piece of handwriting which can be recognized if the viewer holds it at an angle . We also found that this kind of field of view could be recognized easily by defocusing a television image of a system shown in Fig . 7 . In Fig. 8, we show a low scale resolution sampling of a field of view using a fiber optic digitization procedure in which the low level of digitization and interrogation per profile is adequate because the analyst is not interested in the structural details of any individual profile but in the distribution function of the overall features of the field of view . It is envisaged that this kind of low level resolution examination of profile may be particularly appropriate for the Fine Particle Analyst .

1 52

BRLan H . KAYE

+LLLiL111111LLLLLLILLILLLLLI

~111111111L1111It111111111111111111 .1111111999111199911L19991111111111I 1111199777999977799997779911111111' 11111977777887777788777779111111-1 111197777777877777877777791111111 1111977777778777778777777791111111 : 1111977777778777779777777791111111 : 11111977777887777788777779--i11111 11111997779999777999977799111111-11 1111111999111199911119901111111111 : 1111111111111111111111111111111111~ Ullil1111111111ILIi11111111111111

11111111 iiii111111111111].11tt 1L1111111111111111111111111111 .11111100011110001111000111111' 11111009990000999000099900111111111097779999777999977901111' 111109177777977777877777790111' 111109777777877777877777790111 : 111109777777877777877777790-1-1 111110977799997779999777901-11 1 111110099900009990000999001111 1 1111111000111100011110001111111 1111111111111111111111111111111 1111111111111111111111111111111 111111111111111111111111311111' ,,,,

1 I11IILi1.1.1111111111111111111111111 1111100011110001111000111111111111 .1110000000000000000000001111111111 '111009990000999000099900111l111111 -1100977799997779999777900111111111 i1100977777877177917777900111t11111 LI10097779999777999977790011111111L ti110099900009990000499001111111111 tillo00000000000000000000111L111111 L11110001111000IlIL000lllLlIlltllI 1111111111111111111111111111111111 FIG . 6. Schematic representation of successive stages in the strip routine for exploring Lhe

agglomerate profile .

In summary we may point out that in this short review of current research into pattern recognition and the Fine' Particle Analyst, we have tried to emphasize that compared to the general effort proceeding in pattern recognition the Fine Particle Analyst has problems which are relatively simple if the appropriate pre-compression of data can be achieved . Effort should be placed in obtaining an efficient interface from a profile to any subsequent logic operation . It is the hope of the writer that this special issue of l'attern Recognition will stimulate general research into the problems of the Fine Particle Analyst which are relevant to the technology of pattern recognition .



Efficient pattern recognition in fine particle science



•o

1

•* •

• . • •• 0 • • • •



. .

.7.An example of Information collection using low-level digital sampling . Fir

7aw0("00000000 0 . •00•• 0 0uuuwoo0 .aa 00000000000000000000 000000000 ooo00000 0000000 cocoa" io0woo000oo000wo00ao 0000090 P0000000wo000000000 ocac000 000000 0 .0000 NO 0000000 come '00000" 00000 "Ga 00000000 goa0a0o0o000"000w00 0000 00000 0000000000000000000 )000000000000000000000000000 cc 000000000000 000 100000000000000000000000000 0 00000 0000 000000000000000000 000000~a00 ao eo WOw000W000000 00 0000000 0000aoaaowu.Dole 00000000 0 0 OGOw0000 0a"wOGaO 000000000000 a0 as WeWa00 . 00 0 00000000 'W00000000000000000000000aawao000 00000000000 000000000000 00000000000000000ooa 00 ,Do 000000 00 ~000"coo 0000000 a Coco (COO 0000 00000000000 0000000 ¢7p00000000000 0000000 000 •0 000000 000000000ow 00000' 00000000000000 0000000000 00000000000~1Ya00000000 0wooc 00009000 •0 000 000x"0000 oncon"000000000S0000000 owooaouowwacooooaawooow00Yow a 000000 ooNO aw0g~0a~G8 0 888E88a8`8$ 0 a 28E o S USA 000 0000 00 ec . 000000000000010000000000000000000 0000 Go00000" 0o0000ca0ca r ooaocoo0to 0 00000000000 00000000000000000 00000 0000aa00 W 00 000O00ooca0oooa00ac0aoaaoo 000000 Wwoaoooa w oaooHwooo . 0 .0 caw a0a00ao00oue 1 00000000noaacu0 Doc 00000 000 0000000 00000 0000 'Cooo0000GO0o0000000000000 0000 0 000 i 0 0oq~c0 0 00 00 0 0 0000 00000 0C c00 o00ao0 0o0ooa0 0o00o ac: 0 00c00 0000o0~ 000000000 00 0000100COeno oooo0000 000 00000000000000000000000000004000000 0 0000i000CO0 0 0 00owoooo o "( a O00000OcGO0000 0 00000OOO 00DO 00CC GO onto coal DOGWOOD( 00 0 00000000 c ac cc 0000 0 00 i0Daw000 000400000 000" 000000W 00000cc00 0ooc 40000000 000000000 oc0 0000 0000,0000 00000000 Coca 0 0000 00000 00000 0 00000 0 oaaoo caawoac0a000 o0oo 00000 0 000000 occcc00wccc a once o owo0000000aa000000 00000000 00000 0acaooeaooea0 0 0000 0000 0 0 0000 000000 1.00400000006000000 a0noo0oc 0 . 00nwa eaOOa "Deco aa0w0 ca0pc0C 0000000 0 00000000 00 0 000 a00 c00cc o a 00oaao0ooaaco0000 0 a a 00000 0000000000000000000000 Goo 00000000000 coo oo000000000000 0 OCOODO000a a 00000000 0 000 0aeooc on" 0000000000 00 000000DOC0 00000000a"OD a o00oDO000 000'0000 •0 00000G0o0000000000"001 "D 00co0o00mom 0000000000 000000 O000000oc0o 00 00000000000UN 000 coo woo 00w•0a00o"0"0000 00co0 000 OOON000000 0 0000000000COa0 000000 00 00 wa000oa/ta000000 a00 0000 000000090000090000 00 0 coo GO o000co000 aooo auwoca00 wo0000ba000 0 000o"Doow o"Doow •ow 000000000 0000000 00 00 co o 00 co 0 cog "000oo0ooaO00 aWDOao 0000000000 000 Ca0n000a0cu row 000000 oleo 0 00 0 0 000 1 a000cooGo caa0ao"wa0ao 0000000 000000 00c0c 000000000000000 oowoo0000o w000ow000 " oo00o 0000000 wen Dean 0a0 000000000 000000000 on 0000 •000000000 0x10000000000000 00 0o000wo000w o 00000000000000000000 00 ONO" ON . .. . Do w 0ON 0000oaao 000 0 ooa Do o 00000 000000000 00 0000000 0 000000 00 000000100 Ooaoaa 004000000 0000000000000000 0awooa000 0o1Owo00owo000000 0000000t00 o a0aoow *000000000000 OOOa a"00 a000an0000C00ca000000GO0G000 0 000000 000000000 Ow 000000000001000 i 0 0 0000000 9000 000oooa040 ooeuao00 ooaoaa 0000 c oocqg00000 000000 00000000 00 0000000 o0 0 •00000gcow G O Ga 0 ..Coco wo"000w00a0 "o0goa0gcp 000 00000 000 occ .00000000000~00000 0000 00 0 00000WOOUOOw 0000000 oa00 • 0000 00000e00nn00 000 40 000000 0w• t_. -t- aooooo0wooooo00o Deco"0 oao 000* coca 000 •cO0uw00 0 oaoa "0000 0000000000000 roe Data 0 aooo era oao" 000040 own ca

fa; . H . Sample area

of

low resolution digital representation

of the test field of view .

li3



154

Bit a H . KATE

REFERENCES 1 . G . C . Cx ,a, R . S . Lrnj .nY, D. K . POLLOCK and A . Rozn nLD . Pattern Recognition . Thompson Book Co . . . DC . (1968) . Washington 2 . 13i bliography on Automated Microscope Systems with 233 references . 'I he Microscope 19.104-112 (1971) . 3 . See several papers in the Proceedings of Bradford Conference on Particle Size Analysis, Sept., 1970 . Proceedings published by British Society for Analytical Chemistry . 14 Savdle Row . London, Hngland . 4 . B . H . KAYE, A simulated array of particle profiles for use in microscope methods of particle size analysis, Powder Technology 4. 275-27.9 ( 1970171) . 0. V . E . GIULANO, How we find patterns . International Science and Technology, 40-51(1967). 6. H . L. DAVis, Whatever happened to Cybernetics? April 1967 . Scientific Research 68-86. 7. 1. D . REDMAN, Optical aids to pattern recognition . Atomic Weapons Research Establishment (1967) . 8 . J . D . REDMAN and C . D . REID . Holography . optical information processing and fiber optics . J .Brit .Nucl . Energy Soc . 8 . 65-77 (1967) . 9 . A . 1 . MrALIA, Pattern recognition problems in the study of carbon black, with Appendix by G . J. Hcnm nc, Pattern Recognition 4, 255 (7972).