Research On The Texture Classification Under Varying Illumination

Texture is one of the important underlying visual features of image, the image texture analysis is one of the important research directions of computer vision, texture classification is an important component of the texture analysis. And extracting features from image is the key of the texture classification. Texture images have more texture information than others, and texture information is also easily influenced by outside factors, such as illumination. Therefore, extracting the illumination invariant features from image, elimination of the influence of lighting, is a hot research of texture analysis. This paper mainly study and summarize the effect of light on texture, and two kinds programs of illumination invariant feature extraction algorithm are proposed, which is based on Block-matching and 3D filtering and wavelet de-nosing model. Summarized as follows:① Color constancy computation is one of the methods for elimination of the illumination influence on image. Image imaging is not only related to the characteristics of the reflection surface on the scene, and also related to the light source. Firstly, color constancy computation estimates the illumination when the image imaging, and then corrects the image color to the canonical light source through correction models, such as the diagonal model.② Extracting the illumination invariant features of image is another way to deal with the impact of light. Through studying the influence of varying illumination on image, varying illumination not only affects the color of the image, but also the image texture details, because texture image contains rich texture information. Two kinds programs of texture illumination invariant feature extraction algorithm are proposed with the characteristics of BM3 D algorithm. The idea of the first program is combining with the multi-resolution characteristic of wavelet transform, low and high frequency component can be obtained after using wavelet transform on logarithmic domain of each color channel of texture image, and they contain contour and edge information and texture details respectively, in order to improve the robustness of the illumination invariant feature, we dispose the low and high frequency component with different methods. With regard to the low frequency component, a low-pass filter is used to preserve the texture edge. For the high frequency component, we use block-matching and 3D filtering（BM3D） de-noising method, with the excellent de-noising features of BM3 D algorithm, such as good detail preservation and avoiding blocking artifacts etc. And wavelet reconstruction is utilized to extract the illumination invariant, using principal component analysis（PCA） to reduce the feature dimension, obtain the feature vector, and lastly use the K-Nearest Feature Line（K-NFL） classifier to classify image. Experimental results of texture classification based on Outex_TC₀0014 texture database show that this method delivers good classification performance. Finally, the second program is proposed with pre-processing idea and the difference characteristic of coefficient distribution between spatial domain and logarithmic domain of texture image, and the experimental results of texture classification show that the performance of the second program is improved by 6% than the first program.