Research, Content-based Image Retrieval Method

Content-based image retrieval is a very active research area in computer vision, image processing and pattern recognition with many applications such as Multimedia database system, Digital library, Medical diagnosis, Military reconnaissance, Information retrieval service, Web image search and Electronic Commerce etc. There are four main stages in image retrieval including image input, image preprocessing, feature extraction and similarity measurement. The removal of noise is an important and traditional problem of image preprocessing. There are many works on the restoration of images corrupted by noise. Feature extraction and similarity measurement are critical techniques of image retrieval. Textural feature is always the basis of image analysis, and plays an important role in the content-based image retrieval. Similarity measurement is based on a computed distance between the signatures of query image and that of each image in the database. Different measurement method has a great influence on the performance of image retrieval. The main contributions of this dissertation are summarized as follows:(1) In terms of image denoising, the paper proposes a novel image denoising method in Contourlet domain. By using Contourlet transform, the noised image is decomposed into a low frequency subband and a set of multi-scale and multidirectional high frequency subbands. Optimal thresholds and neighboring window sizes for each subband are determined by minimizing the loss expectation of estimating Contourlet coefficients, and then image denoising is implemented via shrinkage of Contourlet coefficients. The superiority of the proposed method is the high ability of denoising and preserving texture edges and details of images corrupted by Gaussian noise. It is also shown that the proposed method yields better visual effect and higher the peak signal-to-noise ratio as a result of considering dependencies of Contourlet neighborhood coefficients.(2) Some of the most popular texture extraction methods for retrieval are based on filtering or wavelet-like approaches. A novel content-based image retrieval method using moment features in Contourlet domain is proposed. By using Contourlet transform, the image is decomposed into a low frequency subband and a set of multiscale and multidirectional high frequency subbands. In order to improve its degree of noise tolerance, a threshold filtering is applied to the results of Contourlet transform, and then the geometry moments of the Contourlet coefficient at different scales and directions are obtained as the features of image. Similarity measurement is implemented using Euclidean distance between the features of query image and that of each image in the image database. Compared with the recent methods, this approach makes use of the spatial distribution characteristics of contourlet coefficients of images, and does not need any hypothesis. Experimental results show the superiority of the proposed method in terms of the average rate of retrieving relevant images. It is also shown that this method is relatively robust in the presence of white noise.(3) Similarity measurement is one of the critical techniques of image retrieval. Many current retrieval systems take a simple approach by using typically norm-based distances (e.g. Euclidean distance, Manhattan distance) on the extracted feature set as a similarity function. In General, a variety of statistical models are employed to represent texture images in the wavelet domain, and the common assumption that all prior probabilities of the hypotheses are equal.In fact, the statistical prior probability of the subband coefficients doesn't conform to the average distribution. To tackle this problem, a novel texture image retrieval approach based on the weighted Kullback-Leibler divergence of the principal direction of the texture in the Non-Decimated wavelet domain is presented. By using Non-Decimated wavelet transform, the image is decomposed into a set of multiscale and multidirectional high frequency subbands. A statistical-model based on Gassian Mixture Model is employed and the model parameters are used to form feature vectors for image retrieval. Since the radon transform can detect the principal direction of the texture, employing the probability weight gains the weight of three directions of wavelet decomposition. When the weight is larger, the direction of textures is more appropinquity. Therefore, similarity measurement is implemented via computing the weighted Kullback-Leibler divergence (KLD) to considering the real distribution of prior probability. The experimental results indicate that this method significantly improves retrieval accuracy rates.