Research On Content-Based Image Retrieval Technology

As a kind of important visual informative media, images have been applied to almost all fields of science and technology and all aspects in daily life. Therefore, it is critical for people to retrieve valuable information quickly and efficiently from a mass of image resources as their rapid growth. However, there were many limitations in the traditional Text-Based Image Retrieval(TBIR) technology. In 1990s, Content-Based Image Retrieval(CBIR) technology emerged as the times required, which inaugurated a more intuitionistic and accurate way to search for images through their visual contents. At present, CBIR has become a very hot research topic in the field of intelligent information processing.There are two pivotal technologies in CBIR: the first is to select appropriate image features and to find an effective feature extraction method, and the second is to adopt an accurate and reliable similarity matching algorithm. Focusing on these two aspects, this paper is mainly concerned with the application of a kind of distinctive and robust local invariant feature i.e. SIFT feature and the corresponding fast indexing and robust similarity matching technologies in high-dimensional space for CBIR services.Regarding image feature extraction, SIFT features have great distinctiveness and robustness to describe the local information of an image directly, which is unlike the usually used global features such as color, shape and texture etc. and also different form the local statistical features based on image segmentation.The similarity matching technology in CBIR includes two levels, i.e. feature-level matching and image-level matching. During the phase of feature-level matching, a kind of algorithm of Approximate Nearest Neighbors(ANN) search, combining the KD tree and the Best Bin First(BBF) search algorithm for SIFT feature indexing and matching, is adopted in this paper so as to overcome the problem of"Curse of Dimensionality"due to the 128-dementianal SIFT descriptor. During the phase of image-level matching, a modified voting strategy called the Nearest Neighbor Distance Ratio Scoring (NNDRS) is put forward to calculate a supporting score for each neighbor's corresponding database image as the evidence to retrieve the similar images in image database.To enhance the robustness of matching, RANSAC algorithm is applied as a geometry verification method (Homography Constraint) to double check the primary retrieval results and remove the false matches, which is helpful to improve the precision of the CBIR system.In this paper, an experimental CBIR system based on SIFT features has been implemented. Through the experiments carried out on the ZuBud image databases, it can be concluded that the approach proposed in this paper obtained high recall and precision(over 84%) , and this method is very suitable for such CBIR applications as looking for the almost identical objects or scenes in a image database.