Research On Content-based Image Retrieval

With the rapid development of multimedia technology and the emergence of large amount of images, content-based image retrieval (CBIR) has become one of most important research fields. How to retrieve relevant image from large database rapidly and accurately has become an urgent problem. The traditional CBIR methods have semantic gap between the images'underlying characteristics and high-level semantic. In order to reduce the influence, researchers put forward a lot of improved methods. The most effective ways are local feature expression and relevance feedback.Specific works of this paper mainly include:1. According to the problems of global characteristics can not describe image details accurately and traditional salient points extraction has a fixed threshold, the article puts forward an improved image retrieval based on salient points. The threshold can be set to a given percentage of the sum of all the saliency values. In order to avoid salient points concentrated in dense texture area, we detect 5 * 5 neighborhoods for each salient point and just retain the maximum one if there are salient points. Salient point features are extracted local as well as global ones. The local features utilize space distribution information of salient points. We divide them into three concentric rings and calculate color moments and shape invariant moments in every annular and global texture features. The experimental results show that the algorithm is good for image retrieval.2. In order to overcome shortcomings that the salient points are lack of semantic information, a method for image retrieval based on salient points and SVM relevance feedback is proposed. SVM relevance feedback trains new classification model from learning the relevant and irrelevant samples labeled by the user. Reclassify the feature library, and give the new retrieval results.3. Propose a method for image retrieval based on salient points and fuzzy feature evaluation index. Fuzzy feature evaluation index is computed from the'intraset ambiguity'and the'inerset ambiguity'as obtained from the relevant and irrelevant set of images. Then recalculate weighted Euclidean distance and given the new retrieval results. Experiments show that the performances after feedback retrieval both be improved greatly.4. In order to validate the performance of proposed algorithms, we design and implement a test system for local image database. The system can preview the results through the small grid, choose features, feedback relevant and irrelevant samples, and display the additional information etc.