Sparse Texture Representation Using Affine-Invariant Neighborhoods

This paper proposes a novel texture representation suitable for recognizing images of textured surfaces under a wide range of transformations, including viewpoint changes and non-rigid deformations. Unlike many existing feature extraction methods, which treat the neighborhood of every pixel as a candidate texture element, the proposed algorithm works by selecting a sparse set of affine-invariant local patches. This spatial selection process, besides providing greater computational efficiency and reducing redundancy in texton dictionaries, permits the computation of characteristic scale and neighborhood shape for every texture element. Shape information is used to determine the right support region for computing intensity-based descriptors. When affine invariance is not required, shape can itself become a discriminative feature. The proposed texture representation is evaluated in retrieval and classification tasks using the entire Brodatz database and a collection of photographs of textured surfaces featuring viewpoint changes and non-rigid deformations, as well as variations in illumination and appearance.