- Email: [email protected]
Colour classification of rubberwood boards for fingerjoint manufacturing using a SOM neural network and image processing
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
available at www.sciencedirect.com
journal homepage: www.elsevier.com/locate/compag
Colour classification of rubberwood boards for fingerjoint manufacturing using a SOM neural network and image processing W. Kurdthongmee ∗ Division of Computer Engineering, School of Engineering and Resources Management, Walailak University, Tha-sa-la, Nakorn-si-thammarat 80160, Thailand
a r t i c l e
i n f o
a b s t r a c t
Article history:
In order to produce a high quality rubberwood fingerjoint with highly uniform colour, wood
Received 17 August 2007
boards of naturally different shades and colours are required to be elaborately classified and
Received in revised form
grouped. Within each group, wood boards of comparable shade and colour are then cut and
23 February 2008
joined to form a highly uniform shade and colour fingerjoint of the required dimensions.
Accepted 14 April 2008
Currently, many manufacturers in Thailand still rely heavily on a manual classification process by an expert. In this paper, an automatic approach based on a combination of an image
Keywords:
processing technique and an artificial neural network is presented. The Kohonen self orga-
Automatic classification of wood
nizing map (SOM) is selected and used for training with modified histogram data from the
Self-organizing map
hue colour component of the rubberwood boards’ images. The outcome SOM is then used to
Rubberwood fingerjoint
classify an unknown colour rubberwood board with a novel colour group identification algorithm. The overall approach has proved effective in classifying the unknown colour of boards with as high as 95% accuracy without human intervention. In many cases, the approach provides invaluable information to guide an operator to easily classify the remaining 5%. © 2008 Elsevier B.V. All rights reserved.
1.
Introduction
A fingerjoint is a means of joining pieces of wood together to form a piece of longer length and width. The ends of the wood are formed into a set of interlocking fingers. These fingers are made water resistant with an adhesive coating and are meshed together under pressure. Thailand is a major producer and exporter of raw rubber products, most of Thai’s fingerjoints are made from rubberwoods. Rubberwood (also called parawood in Thailand) is the standard common name for the timber of Hevea brasiliensis. There is one more important feature of rubberwood that is very important in today’s world. Rubberwood is the most ecologically “friendly” lumber used in today’s furniture industry. After the economic life of
∗
the rubber tree, which is generally 26–30 years, the latex yields become extremely low, and the planters then fell the rubber trees and plant new ones. Therefore, unlike other woods that are cut down for the sole purpose of producing furniture, a rubberwood is used only after it completes its latex producing cycle and dies. This wood is therefore eco-friendly in the sense that we are now using what was going to waste. In fact, rubberwood is one of the most durable lumbers used in the manufacturing of today’s home furnishings. As a member of the maple family, a rubberwood has a dense grain character that is easily controlled in the kiln drying process. A rubberwood has very little shrinkage making it one of the more stable construction materials available for furniture manufacturing.
Tel.: +66 7567 2341; fax: +66 7567 2399. E-mail addresses: [email protected], lek [email protected]. 0168-1699/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.compag.2008.04.002
86
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
Fig. 1 – The block diagram of the automatic approach for classification of rubberwood fingerjoints based on a combination of SOM neural network and image processing techniques.
In order to manufacture a high quality rubberwood fingerjoint, a pre-processing stage needs to be performed on the raw materials. This is carried out by classifying naturally different shades of rubberwoods. Within a rubberwood board, the appearance could be resulted from an annual ring, it is, however, less important for furniture manufacturers comparing to the overall colour of the board. This comes from the fact that the quality of fingerjoints relies heavily on the matched colour of the boards used to assemble them rather than the textures. There are some kinds of texture in an image of woods that could, however, affect the quality of fingerjoints which are: an internal hole and/or knot. The boards with these kinds of texture need to be eliminated from the process. A group of comparable colour shade rubberwoods without defects is then used as the raw materials in the fingerjoint forming process. According to fingerjoint manufacturers in Thailand, the whole process of colour classification still relies heavily on employing an expert to visually inspect the rubberwood board. The goal of this research is to develop an automatic approach based on a combination of an image processing technique and an artificial neural network to remove this limitation. Although, the research topic concerning the integration of image processing techniques and neural networks to agriculture are very popular, only a few publications have reported on the utilization of these techniques to classify/grade wood products. For example, in research described by Silven and Kauppinen (1994) a computer vision based method is used to detect defects in softwood lumbers. In this work, it was found that an operator can inspect and classify only 3–5 out of 30 types of defects with the rate of 30 boards/min and 75–80% accuracy. By transforming an RGB image into a gray scale which is less computationally intensive and utilizing a knowledge-based segmentation algorithm, the proposed approach can reach greater than 80% correctness. Kauppinen (2000) introduced a colour-based inspection method in parquet slab grading. The approach divides the image into small rectangular regions and calculates colour percentile features from these areas. Classification is performed in two stages:
defect detection and recognition. The recognition results are further used in determining the final grade for the parquet slabs. Later Niskanen et al. (2001a,b) presented a publication on the detection of lumber defects by use of colour and texture features. A non-supervised clustering based approach, or self organizing map (SOM), was used for discriminating defects and sound wood. The neural network was trained with multidimensional feature vectors containing texture and colour cues for small non-overlapping regions in the input image. The experiments reflected a very promising result. Closely related research is also found in Limsakul et al. (2007). In contrast to our approach which relies on utilizing the sample image as a whole, they proposed using a point illumination source (to be specific a light emitting diode (LED) is used) and a detector as a way to capture reflected colour from a rubberwood board. The light source was programmed to generate only the main colours. During the training phase, the captured colours of different rubberwood board were used to train a back propagation neural network. The unknown colour rubberwood board was treated in a similar way but its colour value is the obtained result from the neural network. There are several weaknesses of this approach. First of all, it processes colour data in the RGB colour space. The captured colours are very sensitive to the brightness of the light sources which are fairly difficult to control. This leads to incorrect training and making colour matching difficult. In addition, the approach utilizes a point sampling technique. This could possibly miss colour variations within the rubberwood boards. Finally, the installation of the apparatus to provide input data capturing is quite difficult and less flexible. Our research aims to design a machine vision system to automatically and correctly classify colours of rubberwood boards for Thai furniture industries. The specific goals are: (1) to draw a conclusion about the image properties which are the most suitable for training a neuron network in order to obtain the best colour classification result, and (2) to develop an efficient algorithm to make use of information from a neuron network to correctly classify colours of rubberwood boards. In
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
87
Fig. 2 – The averaged hue histograms from 30 images of 10 different colour rubberwood boards.
the next section, the details of a proposed automatic approach based on a combination of an image processing technique and an artificial neural network are given. This is then followed by presentation of the results and discussion in Section 3. Finally, the paper is concluded in Section 4.
2.
Materials and methods
Fig. 1 illustrates the overall processes of a proposed combination of image processing and neural network based rubberwood colour classification approach. Also, the sample images of 10 possible colour groups of rubberwood boards are shown in the figure. The approach can be classified into two main stages: the training and the unknown colour classification or mapping stages. With respect to the figure, the required series of image processing steps to make an image suitable for training and classifying are shown within the dotted block on the right-hand side. During the training stage, several RGB images of rubberwood boards whose colours are pre-classified by an expert are passed through the process in order to train the neural network. SOM (Dekker, 1994; Honkela, 2007; Kohonen, 1990; Eklund et al., 2003) is a selected neural network utilized in our proposed approach due to its capability of unsupervised learning. After SOM training stage has finished, the resulting SOM along with the proposed algorithm for colour group identification, to be described later, can be used to automatically classify a colour of an unknown rubberwood board. It is noted that the image of an unknown rubberwood board must also be applied by the similar series of image processing steps before presenting to the SOM.
2.1.
An image acquisition
The proposed approach relies on acquiring an RGB image of rubberwood boards. During the course of our experimentations, all of the still images used for training and mapping SOM were acquired using a single coupled-charge device (CCD) colour camera (XC-711, Sony, Japan) with a zoom lens of 12.5–75 mm focal length (C6Z1218, Cosmicar), an image grabbing and processing board (Meter, Matrox Inc., Canada), and a personal computer (Intel Pentium 4 processor 2.4 GHz). The CCD colour camera was fixed at a position of 1 m above rub-
berwood board sample and was set to focus on the surface of the sample. A visible full spectrum light source used during our experimentations was the GE Sunshine Starcoat® T8(F15T8/SUN 6PK) light bulb. It was installed to illuminate from a distance of 3 m away with an angle of 45◦ with respect to the sample. All the ambient light effects were eliminated by enclosing the visual system within a black box. Matrox Imaging Library (MIL 7.5, Matrox Inc., Canada) was linked to the programs to grab RGB colour images of 1280 × 960 pixels. The CCD camera was employed for image acquisition with 4150 lx and F4.0 opening (iris diaphragm). Images were stored in the hard drive from the format of the camera into JPEG file format. Image processing was performed using Matlab 7.0 and Visual Basic 6.0 programming. In a practical use, in order to obtain the same classification result, these similar image acquisition installation and setting parameters must be followed to automate the classification process.
2.2.
An image processing and neural network training
Having obtained a set of input images, the information is then extracted and used similarly both during the training and mapping stages. As an image contains more than necessary information, image processing algorithms are required to remove unrelated information and keep only a limited set of important information. The required algorithms in order are: (1) Convert from RGB to HSV colour space. (2) Create a normalized histogram with N-bin of hue (H)components. (3) Calculate the first and second derivatives for each member of the histogram. The first step is carried out in order to convert the image into a colour space which is less affected by ambient illumination (Abdullah et al., 2006; Gonzalez and Woods, 2006). Only the H-component is then used to create the histogram of the image. A preliminary experiment was performed on a set of sample rubberwood boards to ensure that the input data after conversion can be used for classification. The results are shown in Fig. 2, which clearly indicates that the histograms are quite different among different board colours. At this point, it
88
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
can be observed that the image is converted into N vectors of two dimensions which are the bin position and its corresponding normalized frequency. The normalized frequency is performed in order to make the approach independent of the size of the input image. We also found from our preliminary experimentations that this is insufficient to train the SOM. The third step above is then required to add additional two dimensions to the N vectors. This results in the total dimension of the input vector for the SOM to be 4: bin position, frequency, first derivative of frequency, second derivative of frequency. Adding the latter two dimensions to the vector makes it more characteristic to a particular colour group of images. From now on, the histogram with the inclusive of these properties is called the modified histogram. The comparison results revealed that these N vectors are quite similar among the images of rubberwood boards whose colour and shade are comparable. At this point, it should be clear that the inputs which are used to train the SOM are vectors with four dimensions from the images of the rubberwood boards with known shades and colours. Also, the unknown rubberwood board image is also processed in the similar way to that described above during the colour identification stage (to be described shortly). One additional dimension of the vector which is added to the group of known colour and shade data is the colour group. This data is used to label the neural cells after the training stage. For the unknown group, this is what we need to identify with the algorithm to be described in the next section. With respect to the input to our experiments, the 24bit RGB images of rubberwood boards whose sizes are 1280 × 960 pixels were pre-classified by an expert of a rubberwood fingerjoint company into 10 different groups. These are: brown, dark brown, double colours, dark gray, light gray, red/gray, dark pink, light pink, white, dark yellow and light yellow. The images on the left hand side of Fig. 1 illustrate some sample images for each group. It is noted that the numbers within the brackets are the colour codes which are used during the SOM labelling and colour group identification processes. In total, there were 30 images within each colour group. Five images from each group were randomly taken to produce the data for the SOM neural training process. The histograms for all images were set to have an equal size of 10 and 15 bins for the first and second experiments. This was done in order to additionally test if the sizes of the histograms had an impact on the classification results.
2.3.
Training the SOM
During the course of our preliminary experimentations, the evaluation version 1.1 of the Neural Networks Tool (NENET) from a group of researchers at Helsinki University of Technology was mainly used both for training and labelling of neural cells. The tool is based on the Kohonen self organizing maps algorithm described in (Kohonen, 1990). For the main experimentations, the SOM PAK was used instead since it has no restriction on the allowable size of SOMs. To automate the experimentations, MATLAB script was developed to link with the SOM PAK command, running on a DOS environment. Many training parameters with different sizes of SOM were studied and analyzed.
Fig. 3 – The sample SOM resulting from training and labelling with some input data.
During the training stage, a number of maps were created in order to determine suitable parameters. The maps were constantly monitored in terms of both quantization error and visual cluster structure. The hexagonal lattice type was preferred for the visualization of the output map. It was recommended by Kohonen et al. (1996) that the map ought to be rectangular, rather than square, in order to achieve a stable orientation in the data space. Commonly, the x-axis should be about 30% greater than the y-axis, thus forming a rectangular output map. Another recommendation is that the training length of the second part should be at least 500 times the number of network units, in order to reach statistical accuracy, which is measured by the mean square error (MSE) (Kohonen et al., 1996). According to our proposed approach, after training and testing the SOM maps more than 50 configurations, we found that the best result was obtained when the size of SOM was 26 × 26. This does not quite conform to the recommendations given above. The initialization parameters were as follows: neighbourhood function of type bubble, hexagonal topology and random initialization with random seed of 0. The learning rates of training were 0.5 and 0.06 and 3150 and 31,500 iterations for the first and second training phase, respectively. The neighbourhood radius was set to 11 for the first phase and 1 for the second. The resulting SOM map is shown in Fig. 3. It is clear that the neural cells are arranged in the form that many clusters of data, with similar properties, are clearly observed. In addition, many neural cells map more than one colour group code. In such cases, it means that several colour groups of rubberwood board images have approximately equal bin position, frequency, first derivative of frequency and/or second derivative of frequency. This causes difficulty in interpretation by considering the SOM map directly. This is reason why an algorithm to automatically make a decision about the exact colour group for a set of extended histogram data is required. The algorithm is detailed in the next section.
2.4. An algorithm to identify the colour of an unknown board While the processes to train and label the SOM are very similar to the original algorithm (Dekker, 1994; Kohonen, 1990), the
89
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
colour group identification process requires some extensions to the original one. First of all, a “popular list” is introduced to keep track of the matched colour. Its size is equal to the number of colour groups required to classify. Upon initialization, all members of the popular list are reset to zero. This means that no colour is matched to a member of the list at all. During the colour group identification process, the image of an unknown rubberwood board is pre-processed with the same procedures as mentioned in Section 2.2. This also results in N vectors of four dimensions. The member of N vectors is taken individually to query the SOM in order to find the matched neural cell. By definition, the matched neural cell with respect to the Kth-member is the one whose Euclidean distance between these two multidimensional vectors is the shortest. In most cases, the matched neural cell was labelled with more than one colour group code. This is due to the fact that the characteristics of all rubberwood images are almost very similar to each other. The list of colour group codes is then what we expect to retrieve from querying the SOM. The list is elaborately examined in order to update the popular list appropriately. That is to say if the list of colour group codes indicates that with respect to the currently matched neural cell the colour codes 1 and 2 were labelled, the members of the popular list at locations 1 and 2 are then updated by incrementing its member value by 1. After all members of the N vectors have already been processed, the popular list is then examined. The unknown colour to be returned is at the location of the popular list with the highest count. For the case that more than one location of the list gives rise to the same highest count, it means that the colour group identification process fails to make a decision. This case was, however, very rare from our experimentations. The following is the proposed algorithm for the colour group identification process of 10 colour groups. It is noted that the lines with “//” are the comments of the pseudo-code.
This algorithm was coded into a software application by use of MATLAB scripting language. It takes two parameters in order to operate which are: the labelled SOM map (the resulting map from the training stage) and the image of an unknown colour rubberwood board. It returns the matched colour if the popular list contains only one member with the highest count. Otherwise, it returns a list of all indices of the popular list with the highest count. The list can be used to guide an operator to make a final decision about the matched colour. In order to verify the correctness of the proposed colour group identification algorithm, five images of the rubberwood boards for each colour from the set of known colour were used to train the SOM. Later, these five images were then used to query the labelled SOM map by using the algorithm. The result confirmed that the algorithm could classify the colour with up to 98% of correctness. For the rest 2%, the classification results could be used to assist us to make a final decision. This was a very promising verification result.
3.
Results and discussion
From our experiments, the results indicated that the proposed approach gave rise to as high as 95% of correctness without human intervention. This occurs when training the SOM with 270 images of known colour rubberwood boards. For the remaining 5%, although the approach failed to automatically identify, it provided invaluable information to guide an operator to correctly classify. The sample case of this occurs when the input image was of “dark yellow” colour. The popular list of the colour group identification algorithm had the following member values: popular list = [ 7
3
2
4
3
6 8
6
8 3]
90
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
It indicated that two members of the popular list had equally highest count of 8. These two members were mapped to light pink and dark yellow colour (see Fig. 1 for the colours and their corresponding indices), respectively. As the algorithm could not make a decision correctly, it, therefore, opened an opportunity for an operator to choose the correct colour between these two values. Following is another sample case with three equal maximum count members. This list was the result of feeding the input image of “light yellow” colour into the identification algorithm: popular list = [ 5
4
4
5
2
9
9
5
4
9]
This indicated that the popular list had found three possible mapped colours which are dark pink, light pink and light yellow colour, respectively. The experimental results also show that the size of the modified histogram does not affect the classification result. However, the sizes both during SOM training and colour classification stages must be the same in order to get the best classification results. To further validate the correctness of the proposed classification model, a K-fold cross validation with K = 5 was performed in our additional experiments. In the experiments, the additional 100 rubberwood boards sampled from another furniture factory were added to the input data set. This increased the total number of sample rubberwood boards to be 370. The images taken from these boards were randomly assigned to five different groups for validating. This was done to guarantee that each group had a mixed data between the boards which had successfully been used in our previous experiment and the new boards. The five-fold cross validation procedures were followed. This was done by using four out of five groups of images for SOM training following the previously described procedures and the rest single group for testing. Table 1 summarised the prediction accuracies of the five-fold cross validation from the experimental results. It indicated that on average 87% of correctness was obtainable. From our further study within the groups of incorrect classification, it indicated that the results from the colour group identification algorithm were capable of guid-
Table 1 – The prediction accuracies of the five-fold cross validation from the experimental results Index of group of images for testing
Validation results (%) Correct
Incorrect (completely incorrect)
1 2 3 4 5
84 87 81 95 86
16 (7) 13 (9) 19 (10) 5 (2) 14 (6)
Average
86.6
13.4
ing us to make the final decisions with 9% of correctness, out of 16%, when the index of group of images for testing was 1 (the first row of Table 1). This means that for this specific group, the approach completely failed to correctly classify as high as 7%. The number within the brackets in the incorrect validation results column indicates the completely incorrect decisions. We also found that these failures resulted from the imperfections of the rubberwood boards; i.e. the boards contained some internal holes and/or knots. These imperfections make the image histogram change from the normal ones. We are in the process of designing an algorithm to automatically remove such imperfections from a rubberwood image during the pre-processing stage. This could be accomplished by utilizing the texture within an input image. While the proposed approach works well for this application domain, it still has many inherent drawbacks. The first one comes from an inherent drawback of a SOM algorithm. That is to say it is fairly difficult to assign the parameters; i.e. the number of iteration, the learning rate and even the size of SOM, which train the map and give rise to the best separated clusters. These parameters were selected by trial and error during the course of our experimentations. Even the size of the map that gives rise to the best results in this application domain does not conform to the recommenda-
Fig. 4 – The relationships between the parameters of the SOM maps and percent of classification correctness.
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
tions given by Kohonen et al. (1996). In order to search for the most appropriate parameters, many configurations of SOMs were produced and their parameters were elaborately studied. Fig. 4 illustrates a set of SOMs which were trained with the learning rates of 0.1 and 0.01 and 1000 and 100,000 iterations for the first and second training phase, respectively. The neighbourhood radius was set to 10 for the first phase and 5 for the second. The parameters within brackets of the x-axis are arranged in the following orders: the number of images used during training stage of the SOMs, the number of neural cells in the x-axis, and the number of neural cells in the y-axis of the resulting map. Therefore, (5, 25, 22) means the SOM was trained with five images of rubberwood boards with the map dimension of 25 and 22 in the x- and y-axis, respectively. For the sake of clarity, we avoid showing the percentage of more than five labelled neural cells in Fig. 4. These type of neural cells have the similar trend as 1–4 labelled ones. With respect to the results, it can be concluded that the configurations of SOM which give rise to high percentages of correctness have a small number of non-labelled cells. The percent of neural cell with 1 label does not have any relationship with the classification performance. Further, for neural cells with more than two labels, the classification performance is indirectly proportional to the number of neural cells of these types. This should be a very important guideline, apart from using a trial and error approach, for selection of an appropriate SOM in order to obtain the highest classification performance. In summary, for the colour group identification algorithm to perform correctly, the trained and labelled SOM map must have neither a single labelled colour nor a whole labelled colour. That is to say the labelled colours should be distributed to all neural cells. Another drawback is that the proposed approach does not provide a dynamic feedback mechanism for the incorrect colour classification result to the SOM. Doing so could increase the percentage of correctness at the cost of increasing the overall operational speed. Last of all, the proposed approach lacks a mechanism to indicate that it cannot identify the colour correctly. Only if the approach is not sure about the exact colour, it outputs the list of possible matched colours as shown above. We are still investigating ways to rectify this drawback. As mentioned above, after failures in the proposed approach, experiments to investigate the SOM were performed in order to study the inherent properties of the SOM that have an impact on incorrect classification results. We have found that if many neural cells of a SOM contain many colour codes, the SOM is likely to cause incorrect classification results. This is, therefore, a guideline for preparing a suitable SOM.
4.
Conclusions
In this paper, the automatic approach based on a combination of an image processing technique and an artificial neural network in the category of Kohonen self organizing maps is presented. The SOMs are trained with an extension version of histogram data from the hue colour component of the known colour rubberwood boards’ images. The extended histogram
91
consists of an original histogram (bin number, frequency)pair, with an additional two members which are the first and second derivatives of the histogram. The latter two members make the input data’s histogram more characteristic to the image. The outcome SOM is then used to classify the image of an unknown colour rubberwood board with a novel colour group identification algorithm. The main aim of the proposed identification algorithm is to rectify the problem that most neural cells are mapped by more than one colour codes. The overall approach can effectively classify the sample rubberwood boards with as high as 95% correctness without human intervention. In many test cases, the approach provides invaluable information to guide an operator to ease classifying the remaining 5%.
Acknowledgements The author would like to thank Mr. Nattaporn Pongphanvitul for providing us with a set of rubberwood images. Also, Assoc. Prof. Dr. Chusak Limsakul of the Faculty of Engineering, Prince of Songkla University, was very helpful in inspiring us to investigate this very interesting research topic. Also, the author would like to thank Asst. Prof. Dr. David J. Harding and all anonymous reviewers for their comments in the previous versions of this paper.
references
Abdullah, M.Z., Mohamad-Saleh, J., Fathinul-Syahir, A.S., Mohd-Azemi, B.M.N., 2006. Discrimination and classification of fresh-cut starfruits (Averrhoa carambola L.) using automated machine vision system. Journal of Food Engineering 76, 506–523. Dekker, A.H., 1994. Kohonen Neural Networks for Optimal Colour Quantization, vol. 1. Computation in Neural Systems, Network, pp. 354–367. Eklund, T., Back, B., Vanharanta, H., Visa, A., 2003. Using the self-organizing map as a visualization tool in financial benchmarking. Journal of Information Visualization 2, 171–181. Gonzalez, R.C., Woods, R.E., 2006. Digital Image Processing, second ed. Prentice Hall. Honkela, T., 2007. Self-organizing maps in natural language processing. Ph.D. thesis. Helsinki University of Technology, Espoo, Finland [WWW document] http://www.cis.hut.fi/∼tho/thesis/ (accessed 14 February 2007). Kauppinen, H., 2000. A two-stage defect recognition method for parquet slab grading. In: 15th International Conference on Pattern Recognition, vol. 4, pp. 803–806. Kohonen, T., 1990. The self-organizing map. Proceeding of IEEE 9, 1464–1480. Kohonen, T., Hynninen, J., Kangas, J., Laaksonen, J., 1996. SOM PAK: The Self-Organizing Map Program Package. Helsinki University of Technology. Limsakul, C., Kiatweerasakul, M., Bilya, J., Pongphanvitul, N., 2007. Classification of the rubber wood color for the finger joint product. In: The 4th PSU Conference of Engineering (in Thai). Niskanen, M., Silvn, O., Kauppinen, H., 2001a. Color and texture based wood inspection with non-supervised clustering. In: Proceedings of the 12 Scandinavian Conference on Image Analysis (SCIA2001), pp. 336–342.
92
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
Niskanen, M., Silvn, O., Kauppinen, H., 2001b. Experiments with SOM based inspection of wood. In: International Conference on Quality Control by Artificial Vision (QCAV2001), vol. 2, pp. 311–316.
Silven, O., Kauppinen, H., 1994. Color vision based methodology for grading lumber. In: Proceedings of the 12th IAPR International Conference on Pattern Recognition, vol. 1, pp. 787–790.
available at www.sciencedirect.com
journal homepage: www.elsevier.com/locate/compag
Colour classification of rubberwood boards for fingerjoint manufacturing using a SOM neural network and image processing W. Kurdthongmee ∗ Division of Computer Engineering, School of Engineering and Resources Management, Walailak University, Tha-sa-la, Nakorn-si-thammarat 80160, Thailand
a r t i c l e
i n f o
a b s t r a c t
Article history:
In order to produce a high quality rubberwood fingerjoint with highly uniform colour, wood
Received 17 August 2007
boards of naturally different shades and colours are required to be elaborately classified and
Received in revised form
grouped. Within each group, wood boards of comparable shade and colour are then cut and
23 February 2008
joined to form a highly uniform shade and colour fingerjoint of the required dimensions.
Accepted 14 April 2008
Currently, many manufacturers in Thailand still rely heavily on a manual classification process by an expert. In this paper, an automatic approach based on a combination of an image
Keywords:
processing technique and an artificial neural network is presented. The Kohonen self orga-
Automatic classification of wood
nizing map (SOM) is selected and used for training with modified histogram data from the
Self-organizing map
hue colour component of the rubberwood boards’ images. The outcome SOM is then used to
Rubberwood fingerjoint
classify an unknown colour rubberwood board with a novel colour group identification algorithm. The overall approach has proved effective in classifying the unknown colour of boards with as high as 95% accuracy without human intervention. In many cases, the approach provides invaluable information to guide an operator to easily classify the remaining 5%. © 2008 Elsevier B.V. All rights reserved.
1.
Introduction
A fingerjoint is a means of joining pieces of wood together to form a piece of longer length and width. The ends of the wood are formed into a set of interlocking fingers. These fingers are made water resistant with an adhesive coating and are meshed together under pressure. Thailand is a major producer and exporter of raw rubber products, most of Thai’s fingerjoints are made from rubberwoods. Rubberwood (also called parawood in Thailand) is the standard common name for the timber of Hevea brasiliensis. There is one more important feature of rubberwood that is very important in today’s world. Rubberwood is the most ecologically “friendly” lumber used in today’s furniture industry. After the economic life of
∗
the rubber tree, which is generally 26–30 years, the latex yields become extremely low, and the planters then fell the rubber trees and plant new ones. Therefore, unlike other woods that are cut down for the sole purpose of producing furniture, a rubberwood is used only after it completes its latex producing cycle and dies. This wood is therefore eco-friendly in the sense that we are now using what was going to waste. In fact, rubberwood is one of the most durable lumbers used in the manufacturing of today’s home furnishings. As a member of the maple family, a rubberwood has a dense grain character that is easily controlled in the kiln drying process. A rubberwood has very little shrinkage making it one of the more stable construction materials available for furniture manufacturing.
Tel.: +66 7567 2341; fax: +66 7567 2399. E-mail addresses: [email protected], lek [email protected]. 0168-1699/$ – see front matter © 2008 Elsevier B.V. All rights reserved. doi:10.1016/j.compag.2008.04.002
86
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
Fig. 1 – The block diagram of the automatic approach for classification of rubberwood fingerjoints based on a combination of SOM neural network and image processing techniques.
In order to manufacture a high quality rubberwood fingerjoint, a pre-processing stage needs to be performed on the raw materials. This is carried out by classifying naturally different shades of rubberwoods. Within a rubberwood board, the appearance could be resulted from an annual ring, it is, however, less important for furniture manufacturers comparing to the overall colour of the board. This comes from the fact that the quality of fingerjoints relies heavily on the matched colour of the boards used to assemble them rather than the textures. There are some kinds of texture in an image of woods that could, however, affect the quality of fingerjoints which are: an internal hole and/or knot. The boards with these kinds of texture need to be eliminated from the process. A group of comparable colour shade rubberwoods without defects is then used as the raw materials in the fingerjoint forming process. According to fingerjoint manufacturers in Thailand, the whole process of colour classification still relies heavily on employing an expert to visually inspect the rubberwood board. The goal of this research is to develop an automatic approach based on a combination of an image processing technique and an artificial neural network to remove this limitation. Although, the research topic concerning the integration of image processing techniques and neural networks to agriculture are very popular, only a few publications have reported on the utilization of these techniques to classify/grade wood products. For example, in research described by Silven and Kauppinen (1994) a computer vision based method is used to detect defects in softwood lumbers. In this work, it was found that an operator can inspect and classify only 3–5 out of 30 types of defects with the rate of 30 boards/min and 75–80% accuracy. By transforming an RGB image into a gray scale which is less computationally intensive and utilizing a knowledge-based segmentation algorithm, the proposed approach can reach greater than 80% correctness. Kauppinen (2000) introduced a colour-based inspection method in parquet slab grading. The approach divides the image into small rectangular regions and calculates colour percentile features from these areas. Classification is performed in two stages:
defect detection and recognition. The recognition results are further used in determining the final grade for the parquet slabs. Later Niskanen et al. (2001a,b) presented a publication on the detection of lumber defects by use of colour and texture features. A non-supervised clustering based approach, or self organizing map (SOM), was used for discriminating defects and sound wood. The neural network was trained with multidimensional feature vectors containing texture and colour cues for small non-overlapping regions in the input image. The experiments reflected a very promising result. Closely related research is also found in Limsakul et al. (2007). In contrast to our approach which relies on utilizing the sample image as a whole, they proposed using a point illumination source (to be specific a light emitting diode (LED) is used) and a detector as a way to capture reflected colour from a rubberwood board. The light source was programmed to generate only the main colours. During the training phase, the captured colours of different rubberwood board were used to train a back propagation neural network. The unknown colour rubberwood board was treated in a similar way but its colour value is the obtained result from the neural network. There are several weaknesses of this approach. First of all, it processes colour data in the RGB colour space. The captured colours are very sensitive to the brightness of the light sources which are fairly difficult to control. This leads to incorrect training and making colour matching difficult. In addition, the approach utilizes a point sampling technique. This could possibly miss colour variations within the rubberwood boards. Finally, the installation of the apparatus to provide input data capturing is quite difficult and less flexible. Our research aims to design a machine vision system to automatically and correctly classify colours of rubberwood boards for Thai furniture industries. The specific goals are: (1) to draw a conclusion about the image properties which are the most suitable for training a neuron network in order to obtain the best colour classification result, and (2) to develop an efficient algorithm to make use of information from a neuron network to correctly classify colours of rubberwood boards. In
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
87
Fig. 2 – The averaged hue histograms from 30 images of 10 different colour rubberwood boards.
the next section, the details of a proposed automatic approach based on a combination of an image processing technique and an artificial neural network are given. This is then followed by presentation of the results and discussion in Section 3. Finally, the paper is concluded in Section 4.
2.
Materials and methods
Fig. 1 illustrates the overall processes of a proposed combination of image processing and neural network based rubberwood colour classification approach. Also, the sample images of 10 possible colour groups of rubberwood boards are shown in the figure. The approach can be classified into two main stages: the training and the unknown colour classification or mapping stages. With respect to the figure, the required series of image processing steps to make an image suitable for training and classifying are shown within the dotted block on the right-hand side. During the training stage, several RGB images of rubberwood boards whose colours are pre-classified by an expert are passed through the process in order to train the neural network. SOM (Dekker, 1994; Honkela, 2007; Kohonen, 1990; Eklund et al., 2003) is a selected neural network utilized in our proposed approach due to its capability of unsupervised learning. After SOM training stage has finished, the resulting SOM along with the proposed algorithm for colour group identification, to be described later, can be used to automatically classify a colour of an unknown rubberwood board. It is noted that the image of an unknown rubberwood board must also be applied by the similar series of image processing steps before presenting to the SOM.
2.1.
An image acquisition
The proposed approach relies on acquiring an RGB image of rubberwood boards. During the course of our experimentations, all of the still images used for training and mapping SOM were acquired using a single coupled-charge device (CCD) colour camera (XC-711, Sony, Japan) with a zoom lens of 12.5–75 mm focal length (C6Z1218, Cosmicar), an image grabbing and processing board (Meter, Matrox Inc., Canada), and a personal computer (Intel Pentium 4 processor 2.4 GHz). The CCD colour camera was fixed at a position of 1 m above rub-
berwood board sample and was set to focus on the surface of the sample. A visible full spectrum light source used during our experimentations was the GE Sunshine Starcoat® T8(F15T8/SUN 6PK) light bulb. It was installed to illuminate from a distance of 3 m away with an angle of 45◦ with respect to the sample. All the ambient light effects were eliminated by enclosing the visual system within a black box. Matrox Imaging Library (MIL 7.5, Matrox Inc., Canada) was linked to the programs to grab RGB colour images of 1280 × 960 pixels. The CCD camera was employed for image acquisition with 4150 lx and F4.0 opening (iris diaphragm). Images were stored in the hard drive from the format of the camera into JPEG file format. Image processing was performed using Matlab 7.0 and Visual Basic 6.0 programming. In a practical use, in order to obtain the same classification result, these similar image acquisition installation and setting parameters must be followed to automate the classification process.
2.2.
An image processing and neural network training
Having obtained a set of input images, the information is then extracted and used similarly both during the training and mapping stages. As an image contains more than necessary information, image processing algorithms are required to remove unrelated information and keep only a limited set of important information. The required algorithms in order are: (1) Convert from RGB to HSV colour space. (2) Create a normalized histogram with N-bin of hue (H)components. (3) Calculate the first and second derivatives for each member of the histogram. The first step is carried out in order to convert the image into a colour space which is less affected by ambient illumination (Abdullah et al., 2006; Gonzalez and Woods, 2006). Only the H-component is then used to create the histogram of the image. A preliminary experiment was performed on a set of sample rubberwood boards to ensure that the input data after conversion can be used for classification. The results are shown in Fig. 2, which clearly indicates that the histograms are quite different among different board colours. At this point, it
88
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
can be observed that the image is converted into N vectors of two dimensions which are the bin position and its corresponding normalized frequency. The normalized frequency is performed in order to make the approach independent of the size of the input image. We also found from our preliminary experimentations that this is insufficient to train the SOM. The third step above is then required to add additional two dimensions to the N vectors. This results in the total dimension of the input vector for the SOM to be 4: bin position, frequency, first derivative of frequency, second derivative of frequency. Adding the latter two dimensions to the vector makes it more characteristic to a particular colour group of images. From now on, the histogram with the inclusive of these properties is called the modified histogram. The comparison results revealed that these N vectors are quite similar among the images of rubberwood boards whose colour and shade are comparable. At this point, it should be clear that the inputs which are used to train the SOM are vectors with four dimensions from the images of the rubberwood boards with known shades and colours. Also, the unknown rubberwood board image is also processed in the similar way to that described above during the colour identification stage (to be described shortly). One additional dimension of the vector which is added to the group of known colour and shade data is the colour group. This data is used to label the neural cells after the training stage. For the unknown group, this is what we need to identify with the algorithm to be described in the next section. With respect to the input to our experiments, the 24bit RGB images of rubberwood boards whose sizes are 1280 × 960 pixels were pre-classified by an expert of a rubberwood fingerjoint company into 10 different groups. These are: brown, dark brown, double colours, dark gray, light gray, red/gray, dark pink, light pink, white, dark yellow and light yellow. The images on the left hand side of Fig. 1 illustrate some sample images for each group. It is noted that the numbers within the brackets are the colour codes which are used during the SOM labelling and colour group identification processes. In total, there were 30 images within each colour group. Five images from each group were randomly taken to produce the data for the SOM neural training process. The histograms for all images were set to have an equal size of 10 and 15 bins for the first and second experiments. This was done in order to additionally test if the sizes of the histograms had an impact on the classification results.
2.3.
Training the SOM
During the course of our preliminary experimentations, the evaluation version 1.1 of the Neural Networks Tool (NENET) from a group of researchers at Helsinki University of Technology was mainly used both for training and labelling of neural cells. The tool is based on the Kohonen self organizing maps algorithm described in (Kohonen, 1990). For the main experimentations, the SOM PAK was used instead since it has no restriction on the allowable size of SOMs. To automate the experimentations, MATLAB script was developed to link with the SOM PAK command, running on a DOS environment. Many training parameters with different sizes of SOM were studied and analyzed.
Fig. 3 – The sample SOM resulting from training and labelling with some input data.
During the training stage, a number of maps were created in order to determine suitable parameters. The maps were constantly monitored in terms of both quantization error and visual cluster structure. The hexagonal lattice type was preferred for the visualization of the output map. It was recommended by Kohonen et al. (1996) that the map ought to be rectangular, rather than square, in order to achieve a stable orientation in the data space. Commonly, the x-axis should be about 30% greater than the y-axis, thus forming a rectangular output map. Another recommendation is that the training length of the second part should be at least 500 times the number of network units, in order to reach statistical accuracy, which is measured by the mean square error (MSE) (Kohonen et al., 1996). According to our proposed approach, after training and testing the SOM maps more than 50 configurations, we found that the best result was obtained when the size of SOM was 26 × 26. This does not quite conform to the recommendations given above. The initialization parameters were as follows: neighbourhood function of type bubble, hexagonal topology and random initialization with random seed of 0. The learning rates of training were 0.5 and 0.06 and 3150 and 31,500 iterations for the first and second training phase, respectively. The neighbourhood radius was set to 11 for the first phase and 1 for the second. The resulting SOM map is shown in Fig. 3. It is clear that the neural cells are arranged in the form that many clusters of data, with similar properties, are clearly observed. In addition, many neural cells map more than one colour group code. In such cases, it means that several colour groups of rubberwood board images have approximately equal bin position, frequency, first derivative of frequency and/or second derivative of frequency. This causes difficulty in interpretation by considering the SOM map directly. This is reason why an algorithm to automatically make a decision about the exact colour group for a set of extended histogram data is required. The algorithm is detailed in the next section.
2.4. An algorithm to identify the colour of an unknown board While the processes to train and label the SOM are very similar to the original algorithm (Dekker, 1994; Kohonen, 1990), the
89
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
colour group identification process requires some extensions to the original one. First of all, a “popular list” is introduced to keep track of the matched colour. Its size is equal to the number of colour groups required to classify. Upon initialization, all members of the popular list are reset to zero. This means that no colour is matched to a member of the list at all. During the colour group identification process, the image of an unknown rubberwood board is pre-processed with the same procedures as mentioned in Section 2.2. This also results in N vectors of four dimensions. The member of N vectors is taken individually to query the SOM in order to find the matched neural cell. By definition, the matched neural cell with respect to the Kth-member is the one whose Euclidean distance between these two multidimensional vectors is the shortest. In most cases, the matched neural cell was labelled with more than one colour group code. This is due to the fact that the characteristics of all rubberwood images are almost very similar to each other. The list of colour group codes is then what we expect to retrieve from querying the SOM. The list is elaborately examined in order to update the popular list appropriately. That is to say if the list of colour group codes indicates that with respect to the currently matched neural cell the colour codes 1 and 2 were labelled, the members of the popular list at locations 1 and 2 are then updated by incrementing its member value by 1. After all members of the N vectors have already been processed, the popular list is then examined. The unknown colour to be returned is at the location of the popular list with the highest count. For the case that more than one location of the list gives rise to the same highest count, it means that the colour group identification process fails to make a decision. This case was, however, very rare from our experimentations. The following is the proposed algorithm for the colour group identification process of 10 colour groups. It is noted that the lines with “//” are the comments of the pseudo-code.
This algorithm was coded into a software application by use of MATLAB scripting language. It takes two parameters in order to operate which are: the labelled SOM map (the resulting map from the training stage) and the image of an unknown colour rubberwood board. It returns the matched colour if the popular list contains only one member with the highest count. Otherwise, it returns a list of all indices of the popular list with the highest count. The list can be used to guide an operator to make a final decision about the matched colour. In order to verify the correctness of the proposed colour group identification algorithm, five images of the rubberwood boards for each colour from the set of known colour were used to train the SOM. Later, these five images were then used to query the labelled SOM map by using the algorithm. The result confirmed that the algorithm could classify the colour with up to 98% of correctness. For the rest 2%, the classification results could be used to assist us to make a final decision. This was a very promising verification result.
3.
Results and discussion
From our experiments, the results indicated that the proposed approach gave rise to as high as 95% of correctness without human intervention. This occurs when training the SOM with 270 images of known colour rubberwood boards. For the remaining 5%, although the approach failed to automatically identify, it provided invaluable information to guide an operator to correctly classify. The sample case of this occurs when the input image was of “dark yellow” colour. The popular list of the colour group identification algorithm had the following member values: popular list = [ 7
3
2
4
3
6 8
6
8 3]
90
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
It indicated that two members of the popular list had equally highest count of 8. These two members were mapped to light pink and dark yellow colour (see Fig. 1 for the colours and their corresponding indices), respectively. As the algorithm could not make a decision correctly, it, therefore, opened an opportunity for an operator to choose the correct colour between these two values. Following is another sample case with three equal maximum count members. This list was the result of feeding the input image of “light yellow” colour into the identification algorithm: popular list = [ 5
4
4
5
2
9
9
5
4
9]
This indicated that the popular list had found three possible mapped colours which are dark pink, light pink and light yellow colour, respectively. The experimental results also show that the size of the modified histogram does not affect the classification result. However, the sizes both during SOM training and colour classification stages must be the same in order to get the best classification results. To further validate the correctness of the proposed classification model, a K-fold cross validation with K = 5 was performed in our additional experiments. In the experiments, the additional 100 rubberwood boards sampled from another furniture factory were added to the input data set. This increased the total number of sample rubberwood boards to be 370. The images taken from these boards were randomly assigned to five different groups for validating. This was done to guarantee that each group had a mixed data between the boards which had successfully been used in our previous experiment and the new boards. The five-fold cross validation procedures were followed. This was done by using four out of five groups of images for SOM training following the previously described procedures and the rest single group for testing. Table 1 summarised the prediction accuracies of the five-fold cross validation from the experimental results. It indicated that on average 87% of correctness was obtainable. From our further study within the groups of incorrect classification, it indicated that the results from the colour group identification algorithm were capable of guid-
Table 1 – The prediction accuracies of the five-fold cross validation from the experimental results Index of group of images for testing
Validation results (%) Correct
Incorrect (completely incorrect)
1 2 3 4 5
84 87 81 95 86
16 (7) 13 (9) 19 (10) 5 (2) 14 (6)
Average
86.6
13.4
ing us to make the final decisions with 9% of correctness, out of 16%, when the index of group of images for testing was 1 (the first row of Table 1). This means that for this specific group, the approach completely failed to correctly classify as high as 7%. The number within the brackets in the incorrect validation results column indicates the completely incorrect decisions. We also found that these failures resulted from the imperfections of the rubberwood boards; i.e. the boards contained some internal holes and/or knots. These imperfections make the image histogram change from the normal ones. We are in the process of designing an algorithm to automatically remove such imperfections from a rubberwood image during the pre-processing stage. This could be accomplished by utilizing the texture within an input image. While the proposed approach works well for this application domain, it still has many inherent drawbacks. The first one comes from an inherent drawback of a SOM algorithm. That is to say it is fairly difficult to assign the parameters; i.e. the number of iteration, the learning rate and even the size of SOM, which train the map and give rise to the best separated clusters. These parameters were selected by trial and error during the course of our experimentations. Even the size of the map that gives rise to the best results in this application domain does not conform to the recommenda-
Fig. 4 – The relationships between the parameters of the SOM maps and percent of classification correctness.
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
tions given by Kohonen et al. (1996). In order to search for the most appropriate parameters, many configurations of SOMs were produced and their parameters were elaborately studied. Fig. 4 illustrates a set of SOMs which were trained with the learning rates of 0.1 and 0.01 and 1000 and 100,000 iterations for the first and second training phase, respectively. The neighbourhood radius was set to 10 for the first phase and 5 for the second. The parameters within brackets of the x-axis are arranged in the following orders: the number of images used during training stage of the SOMs, the number of neural cells in the x-axis, and the number of neural cells in the y-axis of the resulting map. Therefore, (5, 25, 22) means the SOM was trained with five images of rubberwood boards with the map dimension of 25 and 22 in the x- and y-axis, respectively. For the sake of clarity, we avoid showing the percentage of more than five labelled neural cells in Fig. 4. These type of neural cells have the similar trend as 1–4 labelled ones. With respect to the results, it can be concluded that the configurations of SOM which give rise to high percentages of correctness have a small number of non-labelled cells. The percent of neural cell with 1 label does not have any relationship with the classification performance. Further, for neural cells with more than two labels, the classification performance is indirectly proportional to the number of neural cells of these types. This should be a very important guideline, apart from using a trial and error approach, for selection of an appropriate SOM in order to obtain the highest classification performance. In summary, for the colour group identification algorithm to perform correctly, the trained and labelled SOM map must have neither a single labelled colour nor a whole labelled colour. That is to say the labelled colours should be distributed to all neural cells. Another drawback is that the proposed approach does not provide a dynamic feedback mechanism for the incorrect colour classification result to the SOM. Doing so could increase the percentage of correctness at the cost of increasing the overall operational speed. Last of all, the proposed approach lacks a mechanism to indicate that it cannot identify the colour correctly. Only if the approach is not sure about the exact colour, it outputs the list of possible matched colours as shown above. We are still investigating ways to rectify this drawback. As mentioned above, after failures in the proposed approach, experiments to investigate the SOM were performed in order to study the inherent properties of the SOM that have an impact on incorrect classification results. We have found that if many neural cells of a SOM contain many colour codes, the SOM is likely to cause incorrect classification results. This is, therefore, a guideline for preparing a suitable SOM.
4.
Conclusions
In this paper, the automatic approach based on a combination of an image processing technique and an artificial neural network in the category of Kohonen self organizing maps is presented. The SOMs are trained with an extension version of histogram data from the hue colour component of the known colour rubberwood boards’ images. The extended histogram
91
consists of an original histogram (bin number, frequency)pair, with an additional two members which are the first and second derivatives of the histogram. The latter two members make the input data’s histogram more characteristic to the image. The outcome SOM is then used to classify the image of an unknown colour rubberwood board with a novel colour group identification algorithm. The main aim of the proposed identification algorithm is to rectify the problem that most neural cells are mapped by more than one colour codes. The overall approach can effectively classify the sample rubberwood boards with as high as 95% correctness without human intervention. In many test cases, the approach provides invaluable information to guide an operator to ease classifying the remaining 5%.
Acknowledgements The author would like to thank Mr. Nattaporn Pongphanvitul for providing us with a set of rubberwood images. Also, Assoc. Prof. Dr. Chusak Limsakul of the Faculty of Engineering, Prince of Songkla University, was very helpful in inspiring us to investigate this very interesting research topic. Also, the author would like to thank Asst. Prof. Dr. David J. Harding and all anonymous reviewers for their comments in the previous versions of this paper.
references
Abdullah, M.Z., Mohamad-Saleh, J., Fathinul-Syahir, A.S., Mohd-Azemi, B.M.N., 2006. Discrimination and classification of fresh-cut starfruits (Averrhoa carambola L.) using automated machine vision system. Journal of Food Engineering 76, 506–523. Dekker, A.H., 1994. Kohonen Neural Networks for Optimal Colour Quantization, vol. 1. Computation in Neural Systems, Network, pp. 354–367. Eklund, T., Back, B., Vanharanta, H., Visa, A., 2003. Using the self-organizing map as a visualization tool in financial benchmarking. Journal of Information Visualization 2, 171–181. Gonzalez, R.C., Woods, R.E., 2006. Digital Image Processing, second ed. Prentice Hall. Honkela, T., 2007. Self-organizing maps in natural language processing. Ph.D. thesis. Helsinki University of Technology, Espoo, Finland [WWW document] http://www.cis.hut.fi/∼tho/thesis/ (accessed 14 February 2007). Kauppinen, H., 2000. A two-stage defect recognition method for parquet slab grading. In: 15th International Conference on Pattern Recognition, vol. 4, pp. 803–806. Kohonen, T., 1990. The self-organizing map. Proceeding of IEEE 9, 1464–1480. Kohonen, T., Hynninen, J., Kangas, J., Laaksonen, J., 1996. SOM PAK: The Self-Organizing Map Program Package. Helsinki University of Technology. Limsakul, C., Kiatweerasakul, M., Bilya, J., Pongphanvitul, N., 2007. Classification of the rubber wood color for the finger joint product. In: The 4th PSU Conference of Engineering (in Thai). Niskanen, M., Silvn, O., Kauppinen, H., 2001a. Color and texture based wood inspection with non-supervised clustering. In: Proceedings of the 12 Scandinavian Conference on Image Analysis (SCIA2001), pp. 336–342.
92
c o m p u t e r s a n d e l e c t r o n i c s i n a g r i c u l t u r e 6 4 ( 2 0 0 8 ) 85–92
Niskanen, M., Silvn, O., Kauppinen, H., 2001b. Experiments with SOM based inspection of wood. In: International Conference on Quality Control by Artificial Vision (QCAV2001), vol. 2, pp. 311–316.
Silven, O., Kauppinen, H., 1994. Color vision based methodology for grading lumber. In: Proceedings of the 12th IAPR International Conference on Pattern Recognition, vol. 1, pp. 787–790.