The Research And Realization Of Semantic-Based Image Retrieval With The Tool SVM

With the popularization of network and the development of multimedia technology, many images appear in the field of science and technology, people's daily work and life. How to retrieve the images from image database precisely and efficiently, which has been one important issue in the filed of image management and image application. Image retrieval has drawn the eyes of people.Content-Based Image Retrieval (CBIR) extracts visual features as retrieval features, such as image color, texture and shape, etc. For the existence of semantic gap between visual features and people understanding, CBIR can't get good retrieval results. And Semantic-Based Image Retrieval has been the development trend of image retrieval.After summarized image retrieval technology and presented rationale of Support Vector Machine (SVM), this paper proposed a new method for image annotation and Semantic-Based Image Retrieval.The main contributions of this paper are summarized as follows: First, the paper realized image classification with the tool of SVM. For there are several classes of images in image database, the main steps of the algorithm are: training SVMs for each kind of image in classification procedure, constructing the multi-class SVM with the help of multi-class classification strategy and classifying images with the multi-class SVM. The experiment results show that the image classification precision is high.Second, we proposed one new method of semi-automated image annotation. That is to label one part of image manually and then use our algorithm to annotate those unlabelled images automatically. The way is to use the nearest neighbor method and the results of image classification, and to propagate the annotation to the images which will be annotated in database. The results show that the algorithm can annotate images precisely, and the results are satisfied.Finally, this paper realized the function of semantic image retrieval by using keywords. Every image has the keywords which imply the semantic of images after annotation. The essence of searching the images which have one specific semantic is to find out those images that have the keyword corresponding to the semantic. The retrieval results show that our system has high retrieval precision and recall.In the end, the work is summarized and the further research direction is pointed out.