Image Retrieval Algorithm Based On Color And Self-Similar Characteristic

Nowadays, technology develops rapidly, computer and internet have become an indispensable element in people's lives, and people's requirement of information is increasing day by day. Image information with characteristics of visualization, understandability, and large information, has become an important way for users to get information. How to search images among a great deal of images according to people's requirements becomes the hotspot of people's attention gradually, content-based image retrieval with its unique advantages plays a more and more important role in image retrieval field. Content-based image retrieval (CBIR) means that making use of color, shape, texture, semantic characteristics to query images, try to retrieval similar images with the example based on comprehending the content of image. CBIR extracts features and constitutes indexes from images directly, feature extraction and index constitution can be realized by computer automatically, which avoids the subjectivity of manual description, and the workload is decreased greatly. CBIR technology can be applied in image retrieval engine, begrimed images filtration, scene real-time security monitoring, key frame extraction, image retrieval in special fields, and many other fields, which has important practical significance and applicable value.This paper researched and designed feature extraction, dimension decrease, and retrieval algorithms of images through studying some existing image retrieval algorithms, Discrete Cosine Transform (DCT), Singular Value Decomposition (SVD), fractal theory, and referencing a lot of literatures inside and outside country. Experiments showed that good effects had been achieved.First of all, an image retrieval algorithm based on DCT and SVD was proposed. Color feature as the most intuitionistic feature has obvious vision effect. Comparing with geometry features, color feature has definite stability, and quite robust to zoom, translation, and rotation, so color feature attracts most attention in image retrieval field. In order to describe the local color feature of image effectively, an algorithm that doing child-block partition to the image, and making use of DCT to extract feature of every child-block separately was proposed. DCT the hue matrix of each child-block, and extract the former four most important coefficients as the color feature of the child-block. At the same time, in order to describe the color distributing feature of image sufficiently, we divided the image into primary and secondary areas properly according to the photographic art composition principle, which could distinguish the important degrees of different areas. When do the feature matching in image retrieval process, we can endue different weights to different areas, which can give prominence to the primary content of image, and can incarnate the color distributing feature of image sufficiently. Considering overfull feature vector dimension will decrease the retrieval efficiency, an algorithm that making use of SVD to extract feature of each area was proposed, which achieved better effect of dimension decrease. And at the same time a logical distance measure formula was proposed to measure the difference between two retrieval feature vectors. Experiments showed that the proposed algorithm obtained good recall and precision, could incarnate the color distributing feature well, and had good retrieval effect.Secondly, an algorithm based on self-similar characteristic was proposed. Self-similarity characteristic indicates that the thing possesses self-similarity hiberarchy, the most typical delegate of which is the fractal. Because fractal coding can incarnate the self-similarity characteristic of image's structure, it possesses the predominance which other methods never possess. Even then, most fractal methods are all aiming at gray images, and adopted gray transformation to increase compress ratio, so traditional fractal coding method can't incarnate the color self-similarity characteristic well. Illuming by the fractal thinking, a self-similar characteristic coding method adapting color image retrieval was proposed according to the feature of color image, which expanded the coding space from gray space to color space, and could incarnate the color self-similar characteristic effectively. Considering the intuitivism, isometrical division of range block and domain block was adopted, but traditional fractal methods adopting big domain size and small range size; in order to achieve high compress ratio, traditional method adopted gray transformation to find self-similarity, which lost the color self-similarity of image, in this paper, the gray transformation was omitted, which could incarnate the color self-similarity better, and at the same time reduced the calculation; what's more, traditional fractal method adopted absolute coordinate to mark the position of matching blocks, in this paper in order to incarnate the relative position, relative coordinate was adopted to construct coding. At the same time, traditional fractal methods have the disadvantage of large calculation and low velocity, so center diffused algorithm was adopted, which can improve the coding velocity obviously when giving attention to the quality of child-block matching. In order to achieve a better velocity and retrieval effect, feature vector was constructed through singular value decomposing the self-similar characteristic codes, and a similarity formula was defined for supporting the image retrieval. Experiments showed that the proposed algorithms had lower computation, higher velocity, better retrieval effect, and could adapt CBIR of color images well, which were feasible and valid.At last, considering the variety of image features, in order to incarnate image features better, the algorithm based on DCT and SVD and the algorithm based on self-similar characteristic were combined effectively. Firstly, coarse retrieval was carried through color distributing feature, which could ensure certain precision, and then did the retrieval through self-similar characteristic, which could achieve better sorting effect. Experiments showed that this retrieval method achieved better effect than the two methods above, which avoided the localization of using one method separately, and achieved better recall, precision and sorting effect. To sum up, this paper did some shallow researches in image retrieval field, proposed a color distributing feature image retrieval algorithm based on DCT and SVD, an image retrieval algorithm based on self-similarity characteristic, and an image retrieval algorithm based on the combination of color feature and self-similarity characteristic. Along with the development of technology and the progress of society, CBIR technology will also get rid of the stale and bring forth the fresh. I hope that the algorithms proposed in this paper can make slender contribution to the development of CBIR technology.