Performance in classification

To compare the impact of the CLP in front of other approaches of LBP for colour images, we select the basic and complete classification scheme proposed by Arvis. Then we compare CLP to a direct approach combining a texture analysis adding its colour average (GLACI) and cross-channel marginal approach (CCMA) as proposed by Maempa [10]. Following this classification scheme, we develop our results on the two colour texture databases OUTEX (TC 00013, 68 textures), VISTEX (TC0006, 54 textures) as defined by Arvis. However we add ALOT (250 textures, 6 illuminates

and 4 cameras) and STEX (476 textures) databases that include more complex texture images.

Table 5 shows that CLP obtains a higher rate of a good classification for 3 of the 4 evaluated databases. A gain greater than 10% is obtained in the two databases having a more complex spatio-chromaticity (STEX and ALOT) [23].

LBP-GLACI LBP-CCMA CLP Difference

OUTEX 85.7 80.73 82.1 -3.6

VISTEX 97.02 97.45 97.7 0.35

STEX 60.34 71.24 83.9 12.7

ALOT 58.3 70.64 81.3 11.3

Tab. 5 – Good classification rate for grey-level approach with colour information in Local Binary Pattern (GLACI- LBP), Local binary pattern with colour information in cross-channel marginal approaches (LBP-CCMA) and colour local pattern (CLP).

VISTEX database includes some images from the nature, so the inner spatial complexity is more important with multi-scale textures. For such spatial complexity, the local binary pattern approaches are well adapted, with an average gain of 13% in comparison to the OUTEX case. For the OUTEX database, classical LBP approach using the GLACI construction obtains the better results, in accordance to the fact that the colour complexity of OUTEX images is reduced. In this case, there is no necessity to vector processing rather than scalar processing (Gibbs phenomena). 5 Cases with bad classification

Figure 9 shows some images for which the CLP fails in the texture classification task. In the case of ALOT images, the colour complexity is reduced as for OUTEX case, reducing the interest of the CLP feature for this case of texture discrimination. For the images from STEX database, the texture is related to a collection of objects. In this case, the small spatial distance used for the classification (d=3) is not adapted to characterize the texture content.

a) Macaroni (colored) b) Macaroni (penne) c) Chips (natural) d) Shammy Fig. 9– Images from ALOT database that have problems in the classification percentage.

a) Flower 11 b) Food 01 c) Misc. 01 d) Misc. 43

6 Conclusion

Kepting the initial idea of the local binary pattern, we proposed a new expression adapted to colour domain. To obtain a feature coherent to the human vision and allowing to compare texture discrimination obtained by machine and human vision, we selected to base our construction on the colour difference processed in a perceptual colour space. The core of the feature is then the frequency representation of the colour differences sequences for a defined spatial neighbourhood.

Obtained signatures of Colour Local Pattern (CLP) are easy to interpret and helps us to identify trends in changes in texture and colour. Static and homogeneous textures present an important magnitude for the null frequency, while the isotropic textures. For more heterogeneous and complex texture, the energy is transferred into the highest frequency in relation to the ratio between the distance parameter and the texture pattern size.

Colour Local Pattern (CLP) approach is more interesting when the spatio-chromatic complexity of images increases. Basic approaches separating texture and colour are more adapted for OUTEX images. While the performances are better but close from the classical LBP constructions for colour images in the case of the VISTEX database. Finally, for STEX and ALOT images, where the spatio-chromatic complexity is highest, the gain in performance is close from 10%.

Under another point of view, the CLP expression integrating the texture analysis in a local neighbourhood proposed an efficient implementation of the texton notion proposed by Julesz, without the requirement to a segmentation process. Such consideration allow to imagine new psycho-physical experiments to assess the feature performances in texture discrimination.

Our next trends are to develop a full and vector expression for the Colour Local Pattern (CLP) and consequently to embed the angular variation of the vector difference around the circular neighbourhood in the signature construction.

Bibliography

[1] M. Landy and N. Graham, "Visual perception of texture", Visual Neurosciences, pp. 40-52, 2002.

[2] U. Neisser, "Cognitive psychology", Appleton-Century Croft, 1976.

[3] B. Wandell, “Foundations of vision”, Sinauer Associates Inc. Stanford University, 1995.

[4] B. Julesz, "Texture and visual perception", Scientific American, USA, 1965. [5] B. Julesz, "Experiments in the visual perception of texture", Scientific

American, pp. 20-30, 1975.

[6] T. Caelli and B. Julesz, "On perceptual analysers underlying visual texture discrimination", Biological Cibernetics, vol. 1, 1978.

[7] B. Julesz y J. Bergen, “Textons, the foundamentals elements in preattentive vision and perception of textures”, The bell system technical journal, 1979 [8] R. Haralick, S. K. and H. Dinstein, "Textural features for image classification",

[9] B. Poirson and B. Wandell, "Pattern-color separable pathways predict sensitivity to simple colores patterns", Vision Resources, 1996.

[10] T. Maempa and M. Pietikainen, "Classification with color and texture: jointly or separately," ScienceDirect, Pattern Recognition, 2003.

[11] A. Drimbarean and P. Whelan, "Experiments in colour texture analysis", Pattern Recognition Letters, 2001.

[12] V. Arvis, C. Debain, M. Berducat and A. Benassi, "Generalization of the co- occurrence matrix for colour images: Application to colour texture segmentation", Image Anal Estereol, 2004.

[13] C. Palm, "Color Texture classification by integrative co-occurrence matrices", Pattern Recognition, 2004.

[14] A. Martínez Rios, A. Ledoux, N. Richard and C. Fernandez-Maloigne, "Colour Texture clssification and spatio-chromatic complexity," X Conferenza del Colore, Italy, 2014.

[15] A. Martínez Ríos, N. Richard and C. Fernandez-Maloigne, "Alternative to colour feature classification using colour contrast occurrence matrix", QCAV, 2015.

[16] T. Ojala, M. Pietikainen and D. Harwood, "A comparative study of texture measures with classification based on feature distributions", Pattern Recognition, 1996.

[17] A. Porebski, N. Vandenbroucke and D. Hamad, "LBP histogram selection for supervised color texture classification", ICIP, 2013.

[18] L. Nanni, A. Lumini and S. Brahnam, "Local binary patterns variants as texture descriptors for medical image analysis", Artificial Intelligence in Medicine, 2010.

[19] F. Yang, Y.-Y. Xu, S. Wang and h. Shen, "Image-based classification of protein subcellular location patterns in human reproductive tissue by ensemble learning global and local features", Neurocomputing, 2014.

[20] D. Huang, C. Shan, M. Ardabiliam, Y. Wang and L. Chen, "Local Binary Patterns and Its Application to Facial Image Analysis: A Survey", Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, 2011.

[21] M. Pietikäinen, Z. Guoying, A. Hadid and T. Ahonen, “Computer vision using local binary pattern”, London: Springer London, 2011.

[22] A. Porebski, N. Vandenbroucke, L. Macaire and D. Hamad, "LBP histogram selection for supervised color texture classification," ICIP, 2012.

[23] S. Hossain y S. Serikawa, “Texture databases - A comprenhensive survey”, Pattern Recognition Letters, vol. 34, nº 1, pp. 2007-2022, 2013.