A New Method for the Image Texture Representation

Texture is an important element to human vision, and it has been found to provide cues to scene depth and surface orientation. The pity is that it is hard to adequately model texture so far, and most research on texture is carried out on the Brodatz texture collection. However, the appearance of texture in unconstrained imagery is substantially different from that represented in Brodatz set, and textures in these real-world images are distorted by artifacts such as those resulting from non-uniform lighting, shading and warping in the 3-D space. This makes the problem of texture region extraction even more difficult. In view of above reasons, we adopt previous research experience on which texture is useful for decoding shapes and perspective projections, to present the overall process of representing texture using texture histograms, and new representations of texture is based upon histograms of texture elements. The texture elements are defined from the transformation and quantization of QMG (a quadrature mirror filter) wavelet transform energies. Several experimental evaluations of the wavelet texture feature sets that demonstrate excellent performance in classifying Brodatz textures are presented, and several transformations that allow for approximate rotation or scale invariance are also investigated. The texture histogram representation satisfies the objectives of (1) providing a representation for image or regional texture that is compatible with color histograms, (2) being compatible with compressed-domain texture extraction techniques. By using the histogram and binary set representations of texture we efficiently and effectively represent texture information in images and regions. These visual dimensions are extremely useful for image search and retrieval. Finally, we introduce the process of generating the texture histogram and binary texture set-based feature spaces that are symmetric to those for color histograms. The objectives of obtaining these representations of texture are for measuring the similarity of textures and extracting texture regions from images.