Texton-based Texture Classification

Over the last decade, several studies on texture analysis propose to model texture as a probabilistic process that generates small texture patches. In these studies, texture is represented by means of a frequency histogram that measures how often texture patches from a codebook occur in the texture. In the codebook, the texture patches are represented by, e.g., a collection of filter bank responses. The resulting representations are called textons. A recent study claims that textons based on normalized grayvalues outperform textons based on filter responses (such as MR8 filter responses), despite the weaknesses of such imagebased representations for image modelling. The paper investigates this claim by comparing image-based textons with textons obtained using the complex wavelet transform. The complex wavelet transform differs from the MR8 and similar filters in that it employs filters with relatively low support and in that it constructs image representations with less redundancy. Furthermore, the paper investigates to what extent image-based textons are susceptible to 2D rotations of the texture. It compares image-based textons with rotation-invariant textons based on spin images and polar Fourier features. The performance of the new types of textons is evaluated in classification experiments on the CUReT texture dataset. The results of our experiments with the complex wavelet transform support the claim that filter-based textons do not outperform their image-based counterparts. Furthermore, the results of our experiments reveal that image-based textons are very susceptible to 2D rotations of the texture, making image-based textons unapplicable to real-world texture classification problems. We show that strong rotation-invariant texton-based texture classifiers can be constructed by means of textons based on spin images and polar Fourier features.