Image Pattern Classification And Retrieval

This dissertation focuses on image classification and retrieval, involvedwith the applications of Kernel function-based methods and featureextraction methods. The concrete content is as follow:1. Kernel principle component analysis (KPCA) is a new statisticalsignal processing technique which can extract nonlinear features of images.Kernel functions are key elements for improving it's performance. A newkernel principle component analysis method is proposed, which is based onlocal kernel functions associated with the image features, the imageprinciple components extracted by the proposed method can dispict both theglobal and local features of images, and their performances forclassification are better than that extracted by global kernel principlecomponent analysis method.2. As the texture segment is concerned, a segment method is proposed,which is based on the singular value vector of local auto-correlation matrixand support vector machine. The main advantages of the proposed methodinclude that the algorithm is the simpler one for feature extractionapplications, and texture segmentation is easy to do, and the dimensions ofthe feature vector are decrease. Compared with the other methods, theproposed method possess a lower error rate when used in texturesegmentation.3. As the scale and rotation invariant texture image classification isconcerned, a classification method based on Log-Polar transformation andsupport vector machine is realized. The feature vectors are extracted by rowprojection transform to it's corresponding Log-Polar image, support vectormachines are used for the scale and rotation invariant texture imagesclassification based on the extracted feature vectors. The experiment resultsshow that the proposed method is effective.4. Another method for scale and rotation invariant texture imageclassification is realized, which is based on the singular value vector of theauto-correlation matrix of Log-Polar image, the similarity between twoimages is defined as the consine of the angle between corresponding featurevectors . A class center is randomly selected from each class respectively,the maximum liklihood rule is used for classificaton.5. As the binary image retrieval is concerned, the binary image retrievalis realized through using central projection transform. In the proposedmethod, image feature vector is generated by it's central projection data, itis invariant to translation and scaling and rotating. Compared with the othermethods, the proposed algorithm is simpler and easier to be realized.6. Basing on the definition of fractal dimension, the fractal dimensionsof image and the edge and skeleton of image are regard as feature vector.Besides, we define fractal vector which is a novel feature based on thedetailed analysis of the calculation of the box-counting dimension, we alsouse it as one of the image features, image's distance distribution histogramis calculated also. After multiple feature are extracted, an image retrievalmethod is proposed based on multi-query, and we used it in trademarkimage retrieval successfully.