Research on Transductive Support Vector Machine and Its Application in Image Retrieval
|School||Harbin Institute of Technology|
|Course||Computer Science and Technology|
|Keywords||transductive support vector machine image retrieval relevance feedback feature extraction|
Content-based image retrieval (CBIR) is a hot research direction in computer vision and other computer science fields, but the enormous gap between low-level image features and high level semantic concepts hampers the development of CBIR. The relevance feedback based on support vector machine (SVM) is regarded as one of the strategies that can solve this problem effectively. However, the information embedded in unlabeled samples is not exploited in that method. In order to exploit these information sufficiently, we introduced transductive support vector machine (TSVM) into feedback process and studied how to improve TSVM retrieval results. My work is mainly composed by the following three parts:(1)We introduced TSVM into CBIR. At first, comparing the generalization performance with SVM, we prove it is feasible to using TSVM in CBIR. Then analyzing the requirement of image retrieval, we choose non-progressive TSVM approximate algorithm in feedback process. Based on analyzing the characters of feature vector suitable for TSVM, a new color sparse feature is designed and combined with a kind of texture feature. Experiment results show that using either SVM or TSVM learning in feedback process, our mixed feature gets the best retrieval result.(2)We developed active learning and incremental learning can be used in TSVM feeback process to solve some problems when TSVM applied in CBIR, these problems include the retrieval result is not better than SVM learning, consuing time is much more than SVM, and the results are not improved as feedback times increased. Active learing choose the samples which have the lowest confidence in the last learning process, incremental learning assembles all feedback samples in earlier user’s feedback. Taking these in consideration, a TSVM feedback learning algorithm is designed, and experiment results not only show that the proposed method is more discriminative than the feedback process using SVM, but also indicate that TSVM can be well applied in other fields besides text categorization.(3)We degined and developd a TSVM-based image retrieval system. The system integrates all algorithms which are refered in this paper, and it can be not only concerned as an algortithm demo system, but also used as a platform to do experiment and record the result.