A Semi-Supervised Learning Based Relevance Feedback Algorithm in Content-Based Image Retrieval

As a useful solution for address the faultage between image features and semanteme, relevance feedback (RF) became an effective approach to boost image retrieval. In supervised-based machine learning algorithm, insufficient Labeled training data and the unlabeled data in one RF circle can not represent scatter of features space for all irrelevant images, such algorithm used for CBIR did not show a high performance. As a research hot point, semi-supervised, it can utilize unlabeled data to estimate model of RF so that boost the retrieval performance. This paper proposed a new algorithm for RF: make use of expectation maximization (EM) to learn RBF function for RBF neutral network, integrated active learning to void a local value EM learned, and reduce iterations of feedback, as a result this algorithm learned a RF model based on RBF. Experience indicated that: compare to EM and Bayes, efficiency of learner is improved, user's query concept is grasped quickly.