Local Structure Analysis And Classification Methods Of Image Textures

Texture is a basic characteristic of the image surface,which has a strongly visual effect.Hence,in the domains of pattern recognition and computer vision,a majority of researchers are studying the attributes of image textures so that they can extract texture features and then perform texture analysis,such as texture classification,fingerprint recognition,face recognition and so on.It can be found by consulting lots of literatures that methods of representing image textures can be categorized into two classes: special and transform domainbased methods.The former describes textons by analyzing the local structures of image textures,and uses statistics for constructing the distribution of those textons to represent texture images.The advantage of this kind of method is computational simply,easily understanding,and insensitive to image rotation and illumination change.However,the disadvantage is that the macroscopic and the directional and multi-scale information are ignored.The latter represents texture images by constructing energy features of those subbands obtained by decomposing texture images using multi-scale transform tools,such as wavelet and shearlet etc.Its advantage is that the directional and multi-scale information can be used well.However,a large amount of computation is needed for this kind of method,and it is also very sensitive to image rotation.In allusion to those shortages of the above mentioned methods,we have proposed structural difference histogram based texture representation method(SDHR),multiscale counting and difference based texture representation method(MCDR),and multi-scale sampling based texture representation method(MRIR)to use the advantages of existed methods for texture classification.The SDHR based feature contains not only the local texture information,but also the macroscopic contour information of image textures,which mitigates the disadvantages of those conventional special domain-based methods to some extend.The feature extracted based on MCDR can combine multi-scale analysis and local structure analysis efficiently.Compared with those traditional special domain-based methods,its observation scale is even higher,and its computational complexity is relatively lower.The idea of MRIR based method utilizes advantages of the directional and multi-scale information of wavelet and the special domain-based local structure information in constructing a feature of texture images.The MRIR based feature is robust to differentimaging conditions,for example,rotation,scaling,illumination change and so on.In order to verify the effectiveness and discrimination of our proposed three texture representation methods,we conduct many comprehensive experiments on widely used Brodatz,VisTex,CURet,and Outex texture image databases respectively.Experimental results demonstrate that our proposed three texture representation methods can get a satisfactory classification performance.