Texture classification using rotationally blind feature sets from errors of signal approximations
In this thesis, we develop a feature extraction technique for texture classification and Rotation Invariant (RI) texture classification. The scheme requires the use of image decompositions that provide sparse/compressible representations of textures. In this thesis we focus on the use of the Bamberger Pyramid for image decomposition. The Sequential Approximation Error Curve (SAEC) is obtained by iteratively calculating the approximation error between the original pyramid and a thresholded version of the pyramid where the threshold decreases after each operation. We find that SAECs can be used as a texture descriptor with enough differentiating power to be used for texture classification. We use points along the SAEC as feature sets for both texture classification. In the case of RI texture classification, we find these features to be oblivious to rotation, meaning that no further processing to achieve RI features is needed. We identify these features as being rotationally blind. We test the new feature using a Bayes classifier and compare against other well known schemes in the literature using standard test sets in the field. We present comparisons with other image decompositions like the Undecimated Discrete Wavelet transform and the Steerable Pyramid. We conclude that the proposed feature set is not strongly dependent on a particular image representation. ^
Engineering, Electronics and Electrical
Upadhyayula, Surya B, "Texture classification using rotationally blind feature sets from errors of signal approximations" (2007). ETD Collection for University of Texas, El Paso. AAI1445700.