Texture Image Multi-Scale Features Modeling And Retrieval

With the rapid development of multimedia technology and the richness of image data resources,how to quickly and accurately retrieve the required images from the image database is an urgent problem to be solved.In order to solve this problem,the basic theory of wavelet transform was analyzed first,the advantages of wavelet transform for processing one-dimensional signals and the limitations of processing two and multi-dimensional signals were disposed.In order to solve the deficiencies of wavelet transform in the retrieval system,contourlet transform was introduced.Compared with the wavelet,the contourlet transform better the characteristics including multi-scale,localization,low redundancy,multi-direction,shift invariance and so on.It has obvious advantages in image denoising,image compression,texture image retrieval and other fields.After the basic structure of Content-Based Image Retrival(CBIR)based on content was briefly introduced.Various parts of the texture retrieval system was described in detail.In the multi-scale analysis section,starting from the necessity of applying contourlet transform,the Laplacian Pyramid(LP)framework and direction filter(DFB for short)that affects the retrieval characteristics were introduced.The method of extracting the texture feature of the retrieval system,the structure of the feature vector,the feature vector library,the Canberra distance and the similarity function were described in detail.On the other hand,feature vectors,filter types,and decomposition parameters affect the search results.Local Binary Pattern(LBP)mainly extracts image information by considering the shape and texture features of the image,and connects the extracted image information into a feature histogram to accurately represent the image.This method is also often used to construct a texture image retrieval system.Inspired by the LBP idea,the researchers proposed a Local Oriented Similarity Information Booster(LOSIB)to increase the LBP retrieval rate.This operator can effectively improve the retrieval rate of existing systems.However,as an independent texture image description operator,the texture features cannot be accurately characterized.In this paper,improves the LOSIB operator was enhanced to propose a new type of multi-scale texture feature descriptor MSLOSISD(Multi-Scale Local Oriented Similarity Information Statistical Descriptor).After using the LOSIB operator to extract various different direction differences,the new operator performs statistical measurements on these differences to obtain the mean,variance,skewness,and kurtosis of these differences,and cascades these characteristics.The description of the texture image is obtained so that the texture image can be effectively modeled.This new texture feature operator was utilized to construct a texture image retrieval system.The experimental results show that the retrieval system constructed with this operator as the core has a higher retrieval rate,while the dimension of the required feature vectors is lower and the retrieval speed is faster.Contourlet transform needs more complex transformation to process the texture retrieval problem to get the feature vector.The establishment of feature vector library takes a long time,especially when the image database is extended,the feature vector library needs to be updated and upgraded again.These problems affect the efficiency and accuracy of texture image retrieval to some extent.In order to shorten the retrieval time and improve the retrieval speed,the main feature obtained by the contourlet transform was cascaded with the multi-scale local orientation statistical information enhancement operator LOSIB feature to obtain a new feature vector.Experiments show that this feature vector merging contourlet and LOSIB can improve the retrieval rate of texture retrieval system.