a comparative study for texture classification ... - Semantic Scholar
The 5th International Conference on Natural Computation and the 6th International Conference on Fuzzy Systems and Knowledge Discovery (ICNC’09-FSKD’09)
A COMPARATIVE STUDY FOR TEXTURE CLASSIFICATION TECHNIQUES ON WOOD SPECIES RECOGNITION PROBLEM
1Jing
Yi Tou, 1Yong Haur Tay, 2Phooi Yee Lau
1Computer
Vision and Intelligent Systems (CVIS) Group, Universiti Tunku Abdul Rahman (UTAR), Malaysia 2Hanyang University, Republic of Korea E-mails: {toujy,tayyh}@utar.edu.my, [email protected]
OUTLINE
• • • • • • •
Introduction Feature Extraction Classifier Experiments Conclusion Acknowledgement References
Introduction
Introduction • What is wood species recognition? • Wood species recognition is important as different species of wood has different characteristics and features, these might affect the results or products when they are used for different purposes. In the industry, wood material must be examined before it is selected to be used to produce a product. • Applications: construction, conservation, archaeology, forensic investigation, and etc.
Introduction
? Pear Tree (Pyrus calleryana)
Apple Tree (Malus domestica)
? Coconut Palm (Cocos nucifera)
Chinese Fan Palm (Livistona chinensis)
Wood Cross-section Surface
Sepetir (Species A) Sindora coriacea
Punah (Species B) Tetramerista glabra
(Species C) Myristica iners
Merbau (Species D) Intsia palembanica
? Ramin (Species E) Gonystylus bancanus
(Species F) Artocarpus kemando
References: Menon, P. K. B. (revised by Sulaiman, A. and Lim, S. C.). (1993). Structure and identification of Malayan woods: Malayan forest records no. 25. Malaysia: Forest Research Institute of Malaysia.
Texture • Texture is “the variation of data at scales smaller than the scales of interest” • Textures are generally pattern that we can percept from the surface of an object that helps us recognize what it is and to predict the properties that it has, such as a smooth or rough, hard or soft.
References: Petrou, M. and Sevilla, P. G. (2006). Image processing dealing with texture. West Sussex, England: John Wiley & Sons.
Texture • Applicable in many pattern recognition applications: • Wood species recognition, face detection, document analysis, etc.
References: Tuceryan, M. and Jain, A. K. (1998). Texture analysis. In: The Handbook of Pattern Recognition and Computer Vision. 2nd Edition. (pp 207-248). Singapore: World Scientific Publishing.
Wood Species Recognition [Manual] • Wood species recognition is performed on the crosssection surface of the wood samples because it contains most characteristics • Experts uses a hand lens (10 x) to view the surface and identify it with the characteristics observed Cross-Section Surface
Motivation • Introducing potential texture analysis algorithm: • Texture classification is useful to solve various pattern recognition problems • Consistent time for the recognition process: • Experts may take up to days or weeks to determine the species of a wood due to the number of species and difficulties to differentiate certain species • Demand from the market: • Could be used in furniture manufacture, construction, immigration, forensics, wood studying firms and etc
Objectives •
To recognize the wood species through the cross section surface of the wood samples through different texture classification techniques.
•
To compare the performance of each texture classification technique.
Feature Extraction
1. Grey Level Co-occurrence (GLCM) • GLCM shows relationship between grey-scaled pixels values of the image. • Step 1: Select direction and spatial distance to generate the GLCM • Step 2: Generate the GLCM according to the selected parameters
Directions and Spatial Distances
Images
Images
Generated Raw GLCMs Reference: Haralick, R. M., Shanmugam, K. and Dinstein, L. (1973). Textural features for image classification. IEEE Transactions on System, Man, and Cybernectics, 3, 610-621.
1. GLCM (cont.) •
Step 3: Use textural feature functions on the GLCM to extract features
Textural Features
Normalization:
Reference: Haralick, R. M., Shanmugam, K. and Dinstein, L. (1973). Textural features for image classification. IEEE Transactions on System, Man, and Cybernectics, 3, 610-621.
2. Gabor Filters •
Extract features by analysing the frequency domain of the image
•
Step 1: Create Gabor filters according to the radial center frequency and orientation.
Real Part for Gabor filter
Imaginary Part for Gabor filter
Reference: W. H. Yap, M. Khalid, and R. Yusof, “Face Verification with Gabor Representation and Support Vector Machines”, IEEE Proc. of the First Asia International Conference on Modeling and Simulation, 2007, pp. 451-459. c
2. Gabor Filters (cont.) • Step 2: FFT -> convolution -> IFFT Images filtered by the Gabor filters Orientations
Radial Center Frequencies
• Step 3: Principal Component Analysis (PCA) to reduce feature dimension - Singular Value Decomposition (SVD) Reference: W. H. Yap, M. Khalid, and R. Yusof, “Face Verification with Gabor Representation and Support Vector Machines”, IEEE Proc. of the First Asia International Conference on Modeling and Simulation, 2007, pp. 451-459. c
3. Combined GLCM and Gabor Filters •
Texture classification techniques are often combined
•
The GLCM features are combined with the Gabor features by directly appending the Gabor features after the GLCM features. Proven to be better than either GLCM or Gabor filters in some applications.
•
Reference: J.Y. Tou, Y.H. Tay, P.Y. Lau, “Gabor Filters and Grey-level Co-occurrence Matrices in Texture Classification”, MMU International Symposium on Information and Communications Technologies, Petaling Jaya, 2007. J.A.R. Recio, L.A.R. Fernandez, and A. Fernandez-Sarria, “Use of Gabor filters for texture classification of digital images”, 2005.
4. Covariance Matrix •
• •
The covariance matrix is used to represent the co-variance between the Gabor filters It reduces the dimension of the Gabor features The covariance matrices did not lie on Euclidean space, hence Forstners and Moonen’s distance is used
Reference: J. Y. Tou, Y. H. Tay, and P. Y. Lau, “Gabor Filters as Feature Images for Covariance Matrix on Texture Classification Problem”, Lecture Notes in Computer Science - Proc. 15th International Conference on Neural Information Processing, vol. 5507, Nov 2008, pp. 745-751.
Classifier
1. k-Nearest Neighbor (k-NN) • Winning class decided by the k selected best neighbors from the query sample
2. Verification-based Recognition: Feature • GLCM is used as the features • Eight directions are calculated for each test sample rather than rotating the input images
Eight Directions for GLCM Generation
Reference: J.Y. Tou, Y.H. Tay, and P.Y. Lau, “Rotational Invariant Wood Species Recognition through Wood Species Verification”, Proc. 1st Asian Conference on Intelligent Information and Database Systems, Apr. 2009, pp. 115-120.
2. Verification-based Recognition: Verification • Each training samples are used as templates • The energy value is calculated as the distance between all training samples n
E ( x, y ) ( f x(i) f y (i)) 2 i 1
• A threshold value is calculated for each template based on the same species
1 T x x E x x E x 2 • The energy value for the test sample is the minimum energy obtained for the eight directions n
E i 1
min
0...360
( f (i) f (i)) 2
Reference: J.Y. Tou, Y.H. Tay, and P.Y. Lau, “Rotational Invariant Wood Species Recognition through Wood Species Verification”, Proc. 1st Asian Conference on Intelligent Information and Database Systems, Apr. 2009, pp. 115-120.
2. Verification-based Recognition: Recognition • For each test sample, the species with most winning templates accepted is the winning species
Reference: J.Y. Tou, Y.H. Tay, and P.Y. Lau, “Rotational Invariant Wood Species Recognition through Wood Species Verification”, Proc. 1st Asian Conference on Intelligent Information and Database Systems, Apr. 2009, pp. 115-120.
Experiments
Experimental Dataset • Macroscopic wood images from CAIRO (Centre of Artificial Intelligence and Robotics), Universiti Teknologi Malaysia (UTM) • 6 species of wood • 90 training samples for each species • 10 testing samples for each species 1. 2. 3.
4. 5. 6.
Sesendok (Endospermum malaccense) Keledang (Artocarpus kemando) Nyatoh (Palaquium impressinervium) Punah (Tetramerista glabra) Ramin (Gonystylus bancanus) Melunak (Pentace triptera)
Comparison of Different Techniques • Best recognition rate: 85.00% for Gabor-filter-based Covariance Matrix Texture Classification Techniques
Accuracy (%)
GLCM features
76.67
Raw GLCM
78.33
Gabor features
73.33
Combined GLCM and Gabor features
76.67
Gabor filter-based Covariance Matrix
85.00
Verification-based Recognition
78.33
Findings • Strong difference between training and testing samples of sixth species (Melunak, Pentace triptera) Melunak (Training Sample)
Melunak (Testing Sample)
• The train samples (left) has more obvious white lines and black dots but both characteristics are not obvious in the test samples (right)
Findings • Strong similarity between different species Punah (Tetramerista glabra)
Nyatoh (Palaquium impressinervium)
• Samples from Punah, Tetramerista glabra (left) and Nyatoh, Palaquium impressinervium (right) are both sharing similar patterns of black pores and bright markings occurring along the rays.
Findings • Defects causes misclassification in first species (Sesendok, Endospermum malaccense) Defects
• Certain image has a form of defects which are whitish shown in the circled areas
Conclusion
Conclusion • Similar findings compared to general texture classification: • Raw GLCM outperform GLCM features. • Gabor filters produced poorer result. • Combine GLCM and Gabor features outperform both when used independently. • Gabor filter-based covariance matrix performed the best (85%). • Wood species recognition is more challenging that general texture classification due to the natural similarity of textures among wood species.
Future Works • Proposed embedded device for wood species recognition
Acknowledgement
Acknowledgement • The authors would like to thank Y. L. Lew and the Centre for Artificial Intelligence and Robotics (CAIRO) of Universiti Teknologi Malaysia (UTM) for sharing the wood images. • This research is partly funded by Malaysian MOSTI ScienceFund 01-02-11-SF0019.
References
References 1.
Y.L. Lew, Design of an Intelligent Wood Recognition System for the Classification of Tropical Wood Species, M. E. Thesis, Universiti Teknologi Malaysia, 2005. 2. J.Y. Tou, P.Y. Lau, Y.H. Tay, “Computer Vision-based Wood Recognition System”, Proc. Int’l Workshop on Advanced Image Technology, Bangkok, 2007, pp. 197-202. 3. J.Y. Tou, Y.H. Tay, P.Y. Lau, “Gabor Filters and Grey-level Co-occurrence Matrices in Texture Classification”, MMU International Symposium on Information and Communications Technologies, Petaling Jaya, 2007. 4. J.Y. Tou, Y.H. Tay, and P.Y. Lau, “Rotational Invariant Wood Species Recognition through Wood Species Verification”, Proc. 1st Asian Conference on Intelligent Information and Database Systems, Apr. 2009, pp. 115-120. 5. M. Tuceryan, A.K. Jain, “Texture Analysis”, in: C.H. Chen, L.F. Pau, P.S.P Wang (eds.), The Handbook of Pattern Recognition and Computer Vision (2nd Edition), World Scientific Publishing Co, 1998. 6. Y.Q. Chen, Novel Techniques for Image Texture Classification, PhD Thesis, University of Southampton, United Kingdom, 1995. 7. R.M. Haralick, K. Shanmugam, I. Dinstein, “Textural Features for Image Classification”, IEEE Transactions on Systems, Man and Cybernatics, 1973, pp. 610-621. 8. M. Petrou, P.G. Sevilla, Image Processing: Dealing with Texture, Wiley, 2006. 9. J.A.R. Recio, L.A.R. Fernandez, and A. Fernandez-Sarria, “Use of Gabor filters for texture classification of digital images”, 2005. 10. O. Tuzel, F. Porikli, and P. Meer, “Region Covariance: A Fast Descriptor for Detection and Classification”, European Conference on Computer Vision, vol. 1, 2006, pp. 697-704.
Q&A
Thank You!
A COMPARATIVE STUDY FOR TEXTURE CLASSIFICATION TECHNIQUES ON WOOD SPECIES RECOGNITION PROBLEM
1Jing
Yi Tou, 1Yong Haur Tay, 2Phooi Yee Lau
1Computer
Vision and Intelligent Systems (CVIS) Group, Universiti Tunku Abdul Rahman (UTAR), Malaysia 2Hanyang University, Republic of Korea E-mails: {toujy,tayyh}@utar.edu.my, [email protected]
OUTLINE
• • • • • • •
Introduction Feature Extraction Classifier Experiments Conclusion Acknowledgement References
Introduction
Introduction • What is wood species recognition? • Wood species recognition is important as different species of wood has different characteristics and features, these might affect the results or products when they are used for different purposes. In the industry, wood material must be examined before it is selected to be used to produce a product. • Applications: construction, conservation, archaeology, forensic investigation, and etc.
Introduction
? Pear Tree (Pyrus calleryana)
Apple Tree (Malus domestica)
? Coconut Palm (Cocos nucifera)
Chinese Fan Palm (Livistona chinensis)
Wood Cross-section Surface
Sepetir (Species A) Sindora coriacea
Punah (Species B) Tetramerista glabra
(Species C) Myristica iners
Merbau (Species D) Intsia palembanica
? Ramin (Species E) Gonystylus bancanus
(Species F) Artocarpus kemando
References: Menon, P. K. B. (revised by Sulaiman, A. and Lim, S. C.). (1993). Structure and identification of Malayan woods: Malayan forest records no. 25. Malaysia: Forest Research Institute of Malaysia.
Texture • Texture is “the variation of data at scales smaller than the scales of interest” • Textures are generally pattern that we can percept from the surface of an object that helps us recognize what it is and to predict the properties that it has, such as a smooth or rough, hard or soft.
References: Petrou, M. and Sevilla, P. G. (2006). Image processing dealing with texture. West Sussex, England: John Wiley & Sons.
Texture • Applicable in many pattern recognition applications: • Wood species recognition, face detection, document analysis, etc.
References: Tuceryan, M. and Jain, A. K. (1998). Texture analysis. In: The Handbook of Pattern Recognition and Computer Vision. 2nd Edition. (pp 207-248). Singapore: World Scientific Publishing.
Wood Species Recognition [Manual] • Wood species recognition is performed on the crosssection surface of the wood samples because it contains most characteristics • Experts uses a hand lens (10 x) to view the surface and identify it with the characteristics observed Cross-Section Surface
Motivation • Introducing potential texture analysis algorithm: • Texture classification is useful to solve various pattern recognition problems • Consistent time for the recognition process: • Experts may take up to days or weeks to determine the species of a wood due to the number of species and difficulties to differentiate certain species • Demand from the market: • Could be used in furniture manufacture, construction, immigration, forensics, wood studying firms and etc
Objectives •
To recognize the wood species through the cross section surface of the wood samples through different texture classification techniques.
•
To compare the performance of each texture classification technique.
Feature Extraction
1. Grey Level Co-occurrence (GLCM) • GLCM shows relationship between grey-scaled pixels values of the image. • Step 1: Select direction and spatial distance to generate the GLCM • Step 2: Generate the GLCM according to the selected parameters
Directions and Spatial Distances
Images
Images
Generated Raw GLCMs Reference: Haralick, R. M., Shanmugam, K. and Dinstein, L. (1973). Textural features for image classification. IEEE Transactions on System, Man, and Cybernectics, 3, 610-621.
1. GLCM (cont.) •
Step 3: Use textural feature functions on the GLCM to extract features
Textural Features
Normalization:
Reference: Haralick, R. M., Shanmugam, K. and Dinstein, L. (1973). Textural features for image classification. IEEE Transactions on System, Man, and Cybernectics, 3, 610-621.
2. Gabor Filters •
Extract features by analysing the frequency domain of the image
•
Step 1: Create Gabor filters according to the radial center frequency and orientation.
Real Part for Gabor filter
Imaginary Part for Gabor filter
Reference: W. H. Yap, M. Khalid, and R. Yusof, “Face Verification with Gabor Representation and Support Vector Machines”, IEEE Proc. of the First Asia International Conference on Modeling and Simulation, 2007, pp. 451-459. c
2. Gabor Filters (cont.) • Step 2: FFT -> convolution -> IFFT Images filtered by the Gabor filters Orientations
Radial Center Frequencies
• Step 3: Principal Component Analysis (PCA) to reduce feature dimension - Singular Value Decomposition (SVD) Reference: W. H. Yap, M. Khalid, and R. Yusof, “Face Verification with Gabor Representation and Support Vector Machines”, IEEE Proc. of the First Asia International Conference on Modeling and Simulation, 2007, pp. 451-459. c
3. Combined GLCM and Gabor Filters •
Texture classification techniques are often combined
•
The GLCM features are combined with the Gabor features by directly appending the Gabor features after the GLCM features. Proven to be better than either GLCM or Gabor filters in some applications.
•
Reference: J.Y. Tou, Y.H. Tay, P.Y. Lau, “Gabor Filters and Grey-level Co-occurrence Matrices in Texture Classification”, MMU International Symposium on Information and Communications Technologies, Petaling Jaya, 2007. J.A.R. Recio, L.A.R. Fernandez, and A. Fernandez-Sarria, “Use of Gabor filters for texture classification of digital images”, 2005.
4. Covariance Matrix •
• •
The covariance matrix is used to represent the co-variance between the Gabor filters It reduces the dimension of the Gabor features The covariance matrices did not lie on Euclidean space, hence Forstners and Moonen’s distance is used
Reference: J. Y. Tou, Y. H. Tay, and P. Y. Lau, “Gabor Filters as Feature Images for Covariance Matrix on Texture Classification Problem”, Lecture Notes in Computer Science - Proc. 15th International Conference on Neural Information Processing, vol. 5507, Nov 2008, pp. 745-751.
Classifier
1. k-Nearest Neighbor (k-NN) • Winning class decided by the k selected best neighbors from the query sample
2. Verification-based Recognition: Feature • GLCM is used as the features • Eight directions are calculated for each test sample rather than rotating the input images
Eight Directions for GLCM Generation
Reference: J.Y. Tou, Y.H. Tay, and P.Y. Lau, “Rotational Invariant Wood Species Recognition through Wood Species Verification”, Proc. 1st Asian Conference on Intelligent Information and Database Systems, Apr. 2009, pp. 115-120.
2. Verification-based Recognition: Verification • Each training samples are used as templates • The energy value is calculated as the distance between all training samples n
E ( x, y ) ( f x(i) f y (i)) 2 i 1
• A threshold value is calculated for each template based on the same species
1 T x x E x x E x 2 • The energy value for the test sample is the minimum energy obtained for the eight directions n
E i 1
min
0...360
( f (i) f (i)) 2
Reference: J.Y. Tou, Y.H. Tay, and P.Y. Lau, “Rotational Invariant Wood Species Recognition through Wood Species Verification”, Proc. 1st Asian Conference on Intelligent Information and Database Systems, Apr. 2009, pp. 115-120.
2. Verification-based Recognition: Recognition • For each test sample, the species with most winning templates accepted is the winning species
Reference: J.Y. Tou, Y.H. Tay, and P.Y. Lau, “Rotational Invariant Wood Species Recognition through Wood Species Verification”, Proc. 1st Asian Conference on Intelligent Information and Database Systems, Apr. 2009, pp. 115-120.
Experiments
Experimental Dataset • Macroscopic wood images from CAIRO (Centre of Artificial Intelligence and Robotics), Universiti Teknologi Malaysia (UTM) • 6 species of wood • 90 training samples for each species • 10 testing samples for each species 1. 2. 3.
4. 5. 6.
Sesendok (Endospermum malaccense) Keledang (Artocarpus kemando) Nyatoh (Palaquium impressinervium) Punah (Tetramerista glabra) Ramin (Gonystylus bancanus) Melunak (Pentace triptera)
Comparison of Different Techniques • Best recognition rate: 85.00% for Gabor-filter-based Covariance Matrix Texture Classification Techniques
Accuracy (%)
GLCM features
76.67
Raw GLCM
78.33
Gabor features
73.33
Combined GLCM and Gabor features
76.67
Gabor filter-based Covariance Matrix
85.00
Verification-based Recognition
78.33
Findings • Strong difference between training and testing samples of sixth species (Melunak, Pentace triptera) Melunak (Training Sample)
Melunak (Testing Sample)
• The train samples (left) has more obvious white lines and black dots but both characteristics are not obvious in the test samples (right)
Findings • Strong similarity between different species Punah (Tetramerista glabra)
Nyatoh (Palaquium impressinervium)
• Samples from Punah, Tetramerista glabra (left) and Nyatoh, Palaquium impressinervium (right) are both sharing similar patterns of black pores and bright markings occurring along the rays.
Findings • Defects causes misclassification in first species (Sesendok, Endospermum malaccense) Defects
• Certain image has a form of defects which are whitish shown in the circled areas
Conclusion
Conclusion • Similar findings compared to general texture classification: • Raw GLCM outperform GLCM features. • Gabor filters produced poorer result. • Combine GLCM and Gabor features outperform both when used independently. • Gabor filter-based covariance matrix performed the best (85%). • Wood species recognition is more challenging that general texture classification due to the natural similarity of textures among wood species.
Future Works • Proposed embedded device for wood species recognition
Acknowledgement
Acknowledgement • The authors would like to thank Y. L. Lew and the Centre for Artificial Intelligence and Robotics (CAIRO) of Universiti Teknologi Malaysia (UTM) for sharing the wood images. • This research is partly funded by Malaysian MOSTI ScienceFund 01-02-11-SF0019.
References
References 1.
Y.L. Lew, Design of an Intelligent Wood Recognition System for the Classification of Tropical Wood Species, M. E. Thesis, Universiti Teknologi Malaysia, 2005. 2. J.Y. Tou, P.Y. Lau, Y.H. Tay, “Computer Vision-based Wood Recognition System”, Proc. Int’l Workshop on Advanced Image Technology, Bangkok, 2007, pp. 197-202. 3. J.Y. Tou, Y.H. Tay, P.Y. Lau, “Gabor Filters and Grey-level Co-occurrence Matrices in Texture Classification”, MMU International Symposium on Information and Communications Technologies, Petaling Jaya, 2007. 4. J.Y. Tou, Y.H. Tay, and P.Y. Lau, “Rotational Invariant Wood Species Recognition through Wood Species Verification”, Proc. 1st Asian Conference on Intelligent Information and Database Systems, Apr. 2009, pp. 115-120. 5. M. Tuceryan, A.K. Jain, “Texture Analysis”, in: C.H. Chen, L.F. Pau, P.S.P Wang (eds.), The Handbook of Pattern Recognition and Computer Vision (2nd Edition), World Scientific Publishing Co, 1998. 6. Y.Q. Chen, Novel Techniques for Image Texture Classification, PhD Thesis, University of Southampton, United Kingdom, 1995. 7. R.M. Haralick, K. Shanmugam, I. Dinstein, “Textural Features for Image Classification”, IEEE Transactions on Systems, Man and Cybernatics, 1973, pp. 610-621. 8. M. Petrou, P.G. Sevilla, Image Processing: Dealing with Texture, Wiley, 2006. 9. J.A.R. Recio, L.A.R. Fernandez, and A. Fernandez-Sarria, “Use of Gabor filters for texture classification of digital images”, 2005. 10. O. Tuzel, F. Porikli, and P. Meer, “Region Covariance: A Fast Descriptor for Detection and Classification”, European Conference on Computer Vision, vol. 1, 2006, pp. 697-704.
Q&A
Thank You!
