Research On Content-Based Image Retrieval

Due to the steady growth of computer, multimedia, and Internet techniques, a huge amount of images are available. Currently, rapid and effective searching for desired images from large-scale image databases becomes an important and challenging research topic. Content-based image retrieval (CBIR) is the set of techniques to address the problem of retrieving relevant images from an image database based on automatically derived image features. In recent years, CBIR is a very active research direction and has been applied to many fields.In this dissertation, the exploratory research work has been done around the low-level feature extraction, which include color, shape, spatial features and so on. The main contributions of this dissertation are summarized as follows:1. Several key techniques and algorithms of CBIR are deeply analyzed and discussed, such as, color space, the low-level feature descriptions including color, shape, and texture, the similarity measure between the features and the evaluation methods of image retrieval algorithms.2. An image retrieval approach based on color invariant and shape is proposed. Firstly, based on the property of the components of HSV color space, the color invariant, hue, is selected as the color feature of an image. According to H-, S-, and V- component of an image, state matrixes are designed on the base of gray change of the pixels in the image block to describe their shape feature. Then, two methods, one-step transition probability matrix and state correlogram, are introduced to extract shape feature from the state matrixes. The main difference of the two methods is that one-step transition probability matrix reflects the state transformation of the adjacent state and the latter reflects the spatial transformation of the state in the state matrix.3. An image retrieval algorithm based on the color and shape of image region is introduced. Firstly, by deep analysis of the methods on how to specify the main color of image block, a new way is given in this paper. Next, two shape descriptors, flatness and roughness, are defined to extract the shape feature according to the gray change in the image block. Using these two descriptors the blocks of the image are quantified into different types via their statistical character. Finally, we integrate the main hue or color of the image block withthe quantified results of flatness and roughness to retrieval images.4. A novel entropy-based image retrieval (EBIR) method is presented. At first, the shape descriptors that have been used in image retrieval are deeply study. Then, combined with grid descriptor, gray and information entropy, entropy matrix, made up of grid entropy, is adopted as the shape feature of an image. By entropy matrix, the shape feature and gray information of an image are unified to the concept of image entropy. After that, the problems of using entropy matrix as retrieval feature are discussed. In order to solve the problems, two means, the eigenvalue vector and the constant moment vector of entropy matrix, are accepted respectively as the retrieval vector. Finally, the scale invariance, rotation invariance translation invariance and the effect of gray change and noise of the image on the algorithm are studied and proved through experiments. At the same time, the patulous property of information entropy is used to reduce the effect of image noise on the algorithm.5. A new image retrieval algorithm based on color and its spatial distribution is reported. Above all, several improved methods are designed to extract the color spatial information based on the analysis of the means used before. In the light of the new methods and information entropy, color spatial distribution entropy is adopted as the spatial descriptor of color. Therefore, the whole character of each color can be described by the vector which consists of color histogram and color spatial distribution entropy. At the same time, the image entropy and the moment of color spatial distribution entropy are recommended in order to reduce feature dimension. With the combination of color histogram, color spatial distribution entropy, image entropy and the moment of color spatial distribution entropy, different image retrieval schemes are presented.6. Based on the deep analysis of the properties of information entropy and human vision system, several methods, including histogram sorting, histogram area, the integration of histogram sorting and histogram area and the weighted spatial distribution entropy, are given to remove the effect of entropy on the retrieval algorithm and enhance the retrieval efficiency.