Research Of Image Annotation And Retrieval Based On SVM

With the rapid development of internet and multimedia technology, images, videos and audios increase dramatically. How to retrieval useful data rapidly and effectively from tons of information is the hotspot of current multimedia technology research. Recently, content based image retrieval system (CBIR) becomes the main direction of multimedia information retrieval. CBIR extracts low level image features, calculates the similarity of features and chooses the most similar images as the retrieval results. However, the gap between high-level semantics and low-level image features heavily affects the retrieval performance. There are many approaches trying to solve this problem, such as using the semantic information of the objects in images, using machine learning technology to learn the relevance between low level image features and high level semantic concepts, using relevance feedback to learn user's intention etc.Support vector machine (SVM) has been widely used in image classification and image retrieval. Image retrieval can be regarded as a machine learning process, which attempts to train a learner to classify the images in the database as two classes, i.e. positive (relevant) or negative (irrelevant). The basic idea of SVM is to find the exact hyper-plane to separate two classes and maximize the margin between two classes.This paper mainly studies how SVM is used in image annotation and retrieval, and analyzes the merit and weakness of current work. In image annotation, SVM is used to classify images into different semantic classes which can reach image-level annotation and region-level annotation according to the differences of object classes. In image retrieval, SVM is used to combine relevance feedback to improve retrieval precision by gradually modifying the hyper-plane through user's feedback. To improve active learning with SVM and use more unlabeled data, we propose a new algorithm learning three SVMs separately on the color, texture and shape features extracted from labeled images with three different kernel functions. Different algorithms are used in the selection of the disagreement and agreement samples from the unlabeled data and calculation of their confidence degrees. The lowest confident disagreement samples are returned to user to label and added to the training data with the highest confident agreement samples.According to the algorithm that we proposed, we design and develop an image retrieval system based on multiple SVMs and relevance feedback. Experiments show our algorithm can reach a good result in a much less rounds of feedback compared to other algorithms.