Research On Content-Based Image Retrieval

With the rapid development and application of the computer, network and multimedia techniques, the number of digital image database is growing at a shocking speed. How to effectively organize, manage and retrieve the large-scale image database becomes a very important subject in the field of information retrieval. In this research field, image database system plays an important role in the multimedia information system because of its wide employment in many important applications. The manual labeling is labor-consuming, therefore, it is very difficult for traditional information retrieval technology based on key words to achieve an accurate image description and retrieval. The performance can not satisfy the user's request, so we need to find a kind of effective retrieval mode concerning the complex image database. Content-based image retrieval (CBIR) is a set of techniques to solve the problems based on automatically derived image features. In recent years CBIR is a very hot research subject and has been applied in many fields.In this dissertation, lots of research work has been done around some key techniques of CBIR, which include dimensionality reduction of high-dimensional image feature vectors, region of interest location and feature fusion and so on. The present study is the current research focus on image processing and information retrieval. The main contributions of this dissertation are summarized as follows:We proposed the dimensionality reduction algorithm—Locality Preserving Projections(LPP) to remove relation and redundancy of the high dimensional color feature vector of the image. It is linear, but it not only has many linear algorithm merits, such as fine decorrelation, fast calculation and reliable result, but also it takes the nearest neighbor of the image color feature vector into consideration, so that it maintains the original nonlinear topology structure. The algorithm is described as follows: first, it finds k nearest neighbors of every color feature vector; then it constructs the matrix of weights based on the distance between the vector and its neighbor; at last, low dimension vectors are acquired when high-dimensional vectors are projected to low-dimension space though that matrix of weights.During the study of the LPP, we find that the number of the dimension we get after dimensionality reduction affects the retrieval result greatly. If the number is too large, the relevance of high-dimension color feature vectors can't be clearly removed. On the opposite, if the number is too small, the overlap occurs when high-dimension vector projects to low-dimension vector and the original nonlinear topological structure will be destroyed. All these will reduce the precisions of the retrieval.In this dissertation, for resolving the problem of image features dimension reduction in image retrieval based on color features, image features dimension reduction algorithm based on Adaptive Locality Preserving Projection is proposed(ALPP). On the basis of considering the relationship between every color feature vector and its neighbors, by evaluating the effect of Bayesian criteria on image classifying, clustering operation is introduced into dimension-reduction algorithm to determine adaptively the number of dimensions of feature space. It ensures the dimension reduction result both to eliminate the correlation and redundancy among the high dimensional color feature vectors and to preserve the nonlinear topological structure of original data. The findings of the experiment show that due to Corel image database, when the number of returned images is 50, the precision and the recall for ALPP algorithm are higher in retrieval performance than it for PCA algorithm.In CBIR, if the dimension of the image feature is very high, the complexity of building the image feature database and retrieving the similar images will increase seriously. Locally Linear Embedding(LLE) is a algorithm of nonlinear dimensionality reduction. It reconstructs the color feature vectors by its nearest neighbors to describe the global nonlinear structure. Firstly the algorithm finds the k-nearest neighbors of the color feature vector. Then it calculates the matrix of weights with which nearest neighbors of the color feature vector can reconstruct it. Finally, the algorithm computes the low-dimensional embedding vectors which can be reconstructed by the nearest neighbors and the local matrix of weights. LLE takes the local nearest neighbors into consideration, which preserves the nonlinear topological structure of original space after being embedded into low-dimensional space.While reducing the vector dimension with LLE algorithm, the number of nearest neighbors should be determined. In LLE algorithm, the precision of retrieval varies greatly with different number of nearest neighbors. If the number is too small, the embedding will not reflect the nonlinear topological structure. If it is too large, the embedding will lose its nonlinear character, and produce overlaps in low-dimensional space. All of these will influence the precision of the retrieval.In this dissertation we propose a Variable K Neighbors LLE (VKNLLE) method based on the distribution of the original image feature. The VKNLLE method can reduce the color feature vectors dimension with keeping their original nonlinear topological structure into a lower dimension space. The experimental results show that the proposed VKNLLE method can achieve higher precision rate in CBIR.In this dissertation, in view of the traditional content-based image retrieval algorithms, we usually extract the overall features of image, seldom considering image spatial information. But many users may be merely interested in some target in image. In this condition, the image global features will no longer be effective. So we attempt to extract layer feature in the region of interest in color image .This dissertation makes use of the interest point detection to determine the region of interest in color image. Interest point is an important local feature. It collects a lot of important verge information of color image. We adopt Harris algorithm to detect the interest point because there is Gaussian filter in Harris operator in order to filter the noise. The primary work is converting color images to gray scale image before interest point detection, we will do some pretreatment with the number of interest point and distribution in order to determine the region of interest in color image. After the pretreatment of interest point is completed, we can determine the region of interest in color image.In this dissertation a new ROI-Based Multi-features Comprehensive Retrieval algorithm(RBMCR) is proposed. We extract two low-layer features, the color and spatial feature in the region of interest. Firstly we make use of people's sensitive perception in color to nonuniformed quantizing and forming a feature vector as new color feature in the HSV color space. Then we divide the region of interest into blocks and identify the pixel of the biggest interest point value in each sub-block. We make use of the biggest interest point value in each sub-block and the relative positions relationship between the interest point and those in the adjacent sub-block in image to construct a new spatial feature vector. Finally we adopt the weighted sum of two different methods those reflect different feature vectors as similarity measurement.The algorithm we have proposed does feature extraction only in the region of interest, and it will significantly reduce the data processing time. The precision and the recall for the algorithm are higher in retrieval performance. The experimental results show that the proposed RBMCR algorithm can achieve higher precision-recall rate performance, and the retrieval results of the RBMCR algorithm are better than those with single low-level feature. The image retrieval algorithm based on region of interest is superior to those algorithms based on global feature according to the precision-recall rate.