Research On Relevance Feedback In Content-based Image Retrieval

In content-based image retrieval (CBIR), low-level visual feature (color, texture, shape, etc) are extracted to represent the image, and we use already existing image processing techniques to construct a new algorithm to identify the image. In CBIR, we compare the low-level visual of the image to matching image. The retrieval process of such methods is the computer center. There are the semantic gap between low-level visual feature and high-level concept. Relevance feedback can bridge the semantic gap between high-level concepts and low-level visual features. In general, with using a relevant feedback algorithm to retrieval an image, a query point usually achieves to a local optimum after several iterations such that it is very difficult to increase recall and precision. To solve this problem, an adaptive relevance feedback is proposed, which is able to address such a problem mentioned above. In response to these problems, this paper carries out the following research:This paper presents an adaptive feedback algorithm. The main ideal proposed algorithm is that it can automatic detect whether query point arrive at local optimum. If query point reaches optimum, this query point automatic split two query points using modified EM. Two points as original query points, and then reach local optimum respectively. If users don’t satisfy accuracy, further more these two points are spitted into four query points, and achieved local optimum. The process of iteration is not stopped until users satisfy.In this paper a combination of machine learning adaptive feedback algorithm been proposed. The main ideal proposed algorithm is that SVM learning is embedded into the improvement of EM algorithm. According to the response value of the query point, query point is divided into query point of the positive class and negative class query point. By learning the query point of the positive class and negative class, a SVM classifier is constructed and then all queries are classified by SVM classifier. The algorithm can automatic detect whether query point arrive at local optimum. If query point reaches optimum, this query point automatic split two query points using modified EM. Two points as original query points, and then reach local optimum respectively. If users don’t satisfy accuracy, further more these two points are spitted into four query points, and achieved local optimum. The process of iteration is not stopped until users satisfy.