The Research And Development On Content-Based Image Retrieval Technique

With the rapid development of the computer, communication technique,multimedia and network technology, there are more and more resources of digitalimages available. Image retrieval is becoming a hot research area．In the past, themethod of text-based image retrieval was the only way to search images. But thismethod requires manual annotation of images since it’s not feasible to generatedescriptive text automatically. Annotating images manually will take a lot of labor forlarge-scale image database and has the problem of subjective and incomplete. Nowthe Content-Based Image Retrieval technique can solve these problems efficiently. Itextracts visual features as the retrieval features, such as color, texture, shapeetc．Generally, features of the example image needed are extracted firstly, and thencompared with the features of other images in the database, and finally the results areshowed to the user．Extracting features from image is the key issues in it．This thesis analyses and researches the Content-Based Image Retrieval technique,in the base of reading a lot of references published home and abroad. It introducessome Content-Based Image Retrieval systems, mature retrieval algorithm, andsomething about Content-Based Image Retrieval such as characteristic, keytechnology, general framework, system structure, similarity measure technology,retrieval performance evaluation criteria and so on. In extracting and matching colorfeatures, during the study of different color space and color features, we choose HSVcolor model, divide it into non-uniform spaces, and combine three separatedcomponent into a new one-dimensional feature vector. This method reduces thesystem computational complexity and improves the retrieval efficiency. In order tosolve the problem of zero values in traditional histogram, in the base of colorhistogram, this paper introduces an image retrieval algorithm based on accumulativehistogram which improves the retrieval accuracy. In the texture features aspect, westudy the ways to describe texture features, and focus on the image retrievalhistogram based on gray-level co-occurrence matrix. In order to solve the problems inextracting gray co-occurrence matrix in single direction, we put forward an improvedmethod that will extract gray co-occurrence matrix in four directions, and then get theaverage of them, and get higher precision and recall rates. In order to improveretrieval performance of the system, multi-feature for image retrieval is proposed based on the depth studying of single feature. We choose color feature based onaccumulative histogram and texture feature based on gray co-occurrence matrix, andtake the value after normalizing as the new feature vector of image. Comparing withsingle feature, the method based on comprehensive features get higher precision andrecall rates.The above three retrieval methods are used in the image database system, and getthe increases on the efficiency of whole image retrieval, proved by actual application.