Texture analysis as a basic issue in computer vision and image processing has made great progress during the past few decades.Now more and more efforts are being made to invariant texture analysis.However most of the existing methods discussing this problem only focus on translation,rotation or scale transform.Little work is found on affine or perspective invariant texture analysis that is technically more challenging and practically more desirable.This thesis addresses invariant texture analysis.Its focus is on affine invariant texture classification and invariant texture segmentation.The main contributions of this thesis are summarized as follows: 1.A brief survey on existing invariant texture analysis methods is presented. Each approach iS reviewed according to its classification,and its merits and drawbacks are outlined. 2.A new method of affine invariant texture analysis from the viewpoint of structural analysis is proposed.An area ratio map is introduced and then invariant histogram is computed.Extensive experimental results show the efficacy of this method. 3.A novel algorithm of affine invariant texture analysis based on spectral representation is also proposed.The relationship of texture and its spectral representation under affine transform is carefully studied.Texture signatures are derived which can capture texture regularity.Affine invariant texture features are established from these signatures.The performance of this method is demonstrated in invariant texture classification and retrieval. 4.A new invariant texture segmentation method is described based on circular Gabor functions.A new Gabor parameter selection scheme is proposed.We also present a correlation factor to add texture contextual information. Experimental results clearly show that this approach works well in invariant texture segmentation.
修改评论