With the popularity of computer technology and the development of network, the application of multimedia technology, the quantity of digital image database is growing by the astonishing speed, how to organize, to manage and to retrieve the large scale image database effectively is becoming the current which is a popular topic. As the manual labeling depends on subjectivity of individual,and makes a heavy workload. the traditional information retrieval technology that is based on key words is very difficult to achieve a accurate image description and the retrieval, the performance can not satisfy the user's the request gradually, so we need to find a kind of effective retrieval mode in view of the type complex image database.At first, this paper discusses the content-based image retrieval technology presently, as well as the existent problems. At the same time, we also introduce several important main technology of CBIR system design. For example, the characteristics of the content of image, the model of image databases, the extraction and match of feature of image. when introducing the typical image feature extraction and expression, we focus on the color feature, because the color feature has rotation-invariance and scale-invariance, and it is not sensitive to the direction and size of image, so it became the foremost method in CBIR. In this section, the paper cited color histogram, color moments ,color set and such common color of expression, in these expressions , color histogram is used much often than others .In this paper, in view of the traditional content-based retrieval methods, we usually extract the overall characteristics of image, less consideration in image spatial information. But some users may be merely interested in part of image or some target in image, at this time, the image global features will no longer be effective. So we consider that extracting layer feature in the target region of image .There are many ways in identifying the target region of a image, such as image segmentation, or we could adopt the MCI(man computer interaction) that is user choose the region where they have interested . But the best method of segmentation is not materializing, the man-machine interaction technology is not mature, so this method identified the object region is going to study. This paper presents the issue that is using the feature point detection to solve this problem. Feature point is an important local feature, and it comes from corner point detection theory. Due to it's rotation-invariant, then it almost keep from the impact of illumination conditions. In the same time, it can be described as edge curve of the maximum curvature point, corners collect a lot of important verge information of image.Among many angle detection algorithms, according to the experimental image database, this paper adopt Harris algorithm to operate. Because there is Gaussian filter part in Harris operator in order to smooth the noise, due to computing of color image using Harris operator much more complicated than gray-scale image, so in the design of this Image Retrieval System, the primary work is that converting color images to gray-scale image before characteristic point extraction, and then with the number of feature point and distribution we will do some pretreatment, that is when the quantity of feature point is seldom, it is necessary to reduce the threshold in order to convenience the next step of research; when the extraction feature point around the whole image, then we have to appropriate algorithm remove some feature point make sure that confirm the target region accurately. As the pretreatment of feature point is completed, according to treatment feature point on the location of gray-scale we could acquire feature point position in color images, and we can determine the target region in color image.In feature extraction, we take into account only extract a single low-level characteristics from image assuredly have much better retrieval results than other algorithm and it is refer to some particular images, but for that kind of image database which contain complex types and a wide variety of image, this method appears powerless, the basic reason is that the algorithm based on visual characteristics is main refer to the part attributes, it can not reflect users'will and it do lack of personal perception. Therefore, in this paper, we extract two low-layer features that are the color feature and spatial characteristics in the target region.In this paper, the traditional image quantization algorithm is used to extract the color characteristics of the target region in image. that is in the HSV color space, we make use of people's sensitive perception in color to non-uniform quantization color components and forming a feature vector as color characteristics. For extraction of image spatial feature, the paper first make the target area blocked and then identify the biggest feature point value in each sub-block, this feature point of sub-block and in the adjacent block with the biggest feature point value of the pixel location to describe the spatial relationship characteristics. Finally, in this paper we adopt two different methods that reflect different characteristics of the weighted vector sum as retrieval similarity measure. When users give an example images, the system automatically extracts examples of feature vector image and the characteristics of the vector for comparison, in accordance with the sequence similarity from big to small feedback to the search results to users.When we describes the experimental results and analysis, including annular algorithm and color distribution entropy algorithm, the latter algorithm is the improvement of annular histogram algorithm, two algorithms are based on the global image pixel to statistics histogram for Image Retrieval, although in advance we divided image into several annular which have same center, then we extract color histogram images in each annular, but the image is still refer to all the pixels for computing, but this algorithm we proposed is only in the target region do feature extraction, This will significantly cut the data processing time .At the same time, we can see that the recall–precision rate performance accurately survey that in this paper the method based on the color and spatial characteristics of the characteristics is better than the search results on the characteristics of a low-level search results, compared to based on the overall characteristics of the image retrieval algorithm the local region of the image retrieval algorithm has certain advantages。... |