Local Features In Image Retrieval System

With the development of the digital technology recently, the sorts and quantity of the digital images is also growing rapidly. The digital image is informative and intuitive, so it has been the most important information source. Thus, both how to manage the mass information from digital images and how to capture the information wanted from the mass information are challenges in the digital period. To solve the problems above, the technology of image retrieval is proposed. Now, many image retrieval systems based on content have been developed. However, almost all the systems use the global features only and they can not perform universally and precisely.Local features, as an image feature extraction technology, have risen in recent years. Local features have proven to be very successful in applications such as image matching, image retrieval, image recognition, texture recognition, video data mining, image mosaic, recognition of object categories, and so on. Local features describe the local information of the images. Compared with the global features, the local features are more distinctive, invariable and robust. Thus, the local features are more adaptive when the image has blur background, partial occlusion or illumination changes. Using local features in the image retrieval system is very important for expressing the content of the image and matching the image features rapidly and precisely.The technology of local features contains two parts: local feature detector and local feature descriptor. This paper concludes the previous research on local features and proposes two new methods about local feature technology. Finally, the new methods are applied to a content-based image retrieval system and achieve good results.The main contributions of this paper are: 1. After the research and comparison on the existent local feature detectors, this paper proposed a new saliency detector based on the phase spectrum to filter the local features. This method can effectively filter the background features. Using this method, the local features left can express the image better and achieve more effective computing speed. The calculation of the saliency detector proposed in this paper is simpler than others existent.2. After the research and comparison on the existent local feature descriptors, this paper proposed a fast local feature descriptor to solve the problem of complex calculation in the existent descriptors. This descriptor is constructed based on the framework of SIFT, using the qualitative calculation to replace the quantitative calculation in the SIFT descriptor. This strategy can speed up the calculation of the descriptor. The experiment shows this descriptor is has similar robustness with the SIFT descriptor but faster computing speed.3. A content-based image retrieval system is constructed based on the local features. The methods proposed in this paper are also used. The test result of this system shows methods proposed in this paper can effectively improve the correct rate and response time of the image retrieval system.