Image Retrieval Based On Non-Equality Color Histogram And CTAGD Algorithm

Recently, the amounts of digital images all over the world have been developing strikingly fast with the rapid development of technology of multimedia and computer networks. There are images of several kilomega bytes generated by military and civil applications every day. The digital images include a lot of useful information. However, owing to the out-of-order spreading of these images throughout the world, the information covered in these images cannot be accessed and used efficiently, In order to retrieval images rapidly and accurately, many algorithms of content based image retrieval (CBIR) have been proposed.Firstly, the dissertation gives a brief introduction to the background, application, development and technology on content based image retrieval. Then, the details of each key technology about content based image retrieval have been described in this dissertation. Based on in depth studying on extraction of features and comparability matching algorithm of the colors of images, the dissertation proposes an image retrieval algorithm on the basis of HSI colors accumulative histogram, which is in accordance with visual idiosyncrasy of human beings. First, we use HSI color space to compress color space to 72 kinds of representative colors. Then we design 12 overlapped sub-regions to get color accumulative histogram in each region. Finally, the similarity is calculated based on the accumulative histogram of each region, this method is more simple, flexible and effective to extract the color features which avoid the shortcoming of losing the space distribution information in color histogram.The CTAGD algorithm is also proposed in the dissertation. The main idea is to combine the two features to retrieval image, which are Gaussian density feature extracted in compressed domain and texture feature extracted in pixel domain. In order to get the Gaussian density feature, An image is firstly represented in a polar coordinate, then the geometrical center of the image is acquired and the feature vector is obtained through subsection accumulation of pixel values of eight polar angels (0, 45, 90, 135, 180, 225, 270, 315) . Textual features are obtained mainly through calculating the co-occurrence matrix in pixel domain of four directions.In the calculation of similarity of images, we proposed an image matching algorithm based on cyclic correlation coefficient. In this method , correlation measurement first between images and then production of a pre-selected set of images for distance measurement, In this process, the correlation measurement produces a ascending ordered list, upon which 48 images with maximum correlation values are used to construct this image set. Before the final 12 retrieved images are produced, we calculate the distance between the query image and these 48 images, which are arranged in a descending order.The experimental database consists of 10,000 jpeg images. Compared with some existing image retrieval methods, the experimental results show that the proposed algorithm has made some progress in terms of precision and processing speed.