Design Of Image Retrieval System Based On Content Features

Following text retrieval, image retrieval was a new type of information retrieval. In recent decades, image retrieval had become the focus of retrieval and been a fast develops area. Some Internet search engine and research institutions had launched service of’using images to find images’. Based on stage, image retrieval had three styles:based on text information、based on content features and based on semantic. Because image retrieval based on text information developed earlier, it was mature relatively. Current Internet image retrieval service and most existing image retrieval systems were this kind of retrieval. However, relative to the traditional image retrieval based on text information, image retrieval based on content features was not mature and perfect and still in the stage of research and development. For there was no effective method to extract semantic from image, image retrieval based on semantic was still in the stage of concept.In this paper, after studying the existing image retrieval technology and method, a relatively complete framework of image retrieval system based on content features was summed up. Then according to this framework, an image retrieval system based on content features was realized. Retrieval system realized in this paper included process of image segmentation、feature extraction and matching and calculation of image similarity. In this paper, image segmentation based on the following proposed method:First, determined the number of image blocks depending on peaks number of histogram based on texture and color. Then, according to the number of blocks, used the method of k-means partition, the image was divided into several blocks. At last, an image automatic segmentation was realized. In the stage of extracting image features, texture and color features extracted in first phase and shape features of the image block was used to calculate similarity between two blocks of two images. Because the shape image of same object taken from different perspectives was often different, not only similarity calculated directly between two shapes was used, but also similarity obtained from a new method proposed in this paper was adopted to measure similarity between two shapes. In this method, based on similarities between 2d shape and 3d graphics, similarity between two shapes was calculated indirectly. In detail, first of all, extracted features of the 2d shapes and the 3d shape under different perspectives, and then fitted out shape features used a polynomial of perspectives, then problem became into looking for the best value of polynomial variable to make the difference between polynomial values and the shape feature was smallest. This difference was the similarity between the two shapes. After obtaining similarity between image block, similarity between two images was received by matching blocks of two images individually. At last, according to the similarity between the two images, images in library by descending order of similarity were displayed. Finally, the experiment results revealed that the image retrieval in this paper had a higher recall and precision.