Reasearch On Image Retrieval System Based On Local Features

The large amounts of image data lead to the need of high quality of image retrieval. Lack of multimedia information is not the problem which worries people most, but how to find out the right information faster and more accurate. On one hand, only the research on image retrieval system has been strengthened, can we perfect and solidify the theory of image retrieval system. On the other hand, there are still some low-quality image retrieval results in practice which affect the performance of the whole image retrieval system. So it is urgent and imperative to make more efforts on the research of image retrieval system.In this paper, we research on the key technologies in image retrieval system, especially, the SURF descriptor. The main research work and achievements include:1、SURF has been proven to be one of the state-of-the art feature detector and descriptor, and mainly treats colorful images as gray images. However, color provides valuable information in the object description and recognition tasks. This paper addresses this problem and adds the color information into the scale-and rotation-invariant interest point detector and descriptor, coined C-SURF (Colored Speeded Up Robust Features). The built C-SURF is more robust than the conventional SURF with respect to rotation variations. Moreover, we use112dimensions to describe not only the distribution of Harr-wavelet responses but also the color information within the interest point neighborhood. The evaluation results support the potential of the proposed approach.2、In this paper, we add color information into U-SURF and the evaluation results show that when the rotation angle of the pictures is from0°to20°, CU-SURF has better performance than U-SURF and SURF. However, C-SURF has the best performance among the four methods.During the research of the two key technologies, improvement on SURF has been used as a great analysis tool through the whole process. It is used not only to precisely localize the interest points’ location, but also to describe the interest points in detail. So this method not only improves the image correct matching ratio but also improve the performance of whole image retrieval system.