Technology Research And Method Of Content Based Image Retrieval

With the rapid development of multimedia information technology and the growing popularity of internet, the digital image information has massive growth, image storage and management had been a heavy work. How can we find the image which needed by the user from a huge library of images quickly and accurate has become an urgent problem, therefrom, Content-Based Image Retrieval (CBIR) emerged, and become a popular research topic. In this thesis, we focus on the core issue-based image retrieval research, the main contents include:(1) In order to solve the problem of single visual feature cannot fully describe the image content, we presents a fusion of multiple features of image retrieval algorithm. The algorithm first constructs a Zernike chromaticity distribution moments, and used to extract the image color feature; then the image was decomposed multiscale and multidirectional used by the contourlet transform, and compute each sub-band’s variance and entropy as the texture feature of the image; after that the two features are normalized and determined the corresponding weights; Finally, the similarity calculation method was used to calculate the similarity between images.(2) In order to solve the problem of global features of the image is not very good in describing the semantic content of the image; we used points of interest-based image retrieval algorithms. The algorithm uses the ASIFT operator which was invariance of rotation, translation and scaling changes, and it also has complete affine invariance, to extract the point of interest in the color images, and then apply the global color histogram method to statistics these interest points as the texture feature. Simulation results show that, the ASIFT operator can extract the image corner effectively, edge points and the points around these points, this algorithm can better express image content. This study showed that, combined the histogram with bottom features can get an ideal retrieval results.(3) In order to solve the problem of semantic gap between low-level features and high-level semantic, we employ the Adapted GMM algorithm to reduction the dimension of the features, and we also propose feature weighted SVM algorithm which was consider the sample information. This algorithm firstly use PCA process of the original character space, to remove the correlation of the characteristics and noise; then using the Adapted GMM algorithm to convert the original feature space to a new probabilistic feature space, at the same time, reduce the character space dimension; and then we put forward weighted SVM algorithm which based on the sample information, the new algorithm could match the image feedback process well. Simulation results show that, this algorithm can ensure the ideal retrieval results and also short the execution time.