Semi-supervised Active Learning For Relevance Feedback In Image Retrieval

With the rapid development of multimedia and Internet technology, CBIR (Content-Based Image Retrieval) has attracted considerable public concern. However, the early study on content-based image retrieval focused on how to use the image's color, texture, shape and other features information to determine the similarity between images. This cannot solve the "semantic gap" problem caused by the gap between low-level visual features and high-level semantic concepts. The relevance feedback, considered as an effective way of narrowing the semantic gap, has been introduced into the CBIR area. Statistical learning and machine learning has also been involved in the relevant feedback process to improve image retrieval performance. Today, there still exits small sample problems and data redundancy issues in the learning process. This paper adopts dynamic distance metric and clustering method to form the spectral clustering and explore the semi-supervised active learning after integration of spectral clustering. Specific studies are as follows:(1) Using dynamic distance metric to compute the similarities between images. Specifically, according to the user feedback information during the process of relevance feedback, the system adjusts the weight responsible for different low-level features dynamically. Afterwards, the system computes the similarities between images with the help of the dynamic distance metric. The way reflects the user's "subjective" characteristics and highlights the "advantageous" features to improve the accuracy of clustering.(2) Based on dynamic distance metric, the Normal cut spectral clustering on the samples is adopted in the feedback area (samples with rich information). Then we select the representative samples to be set to handle the data redundancy problem during the active learning process. The classifier transfers from by learning from the original redundancy samples in the same semantic set into by those in totally different semantic set. This improves the whole active learning. (3) Discussing about SVM-AL, SSAIR, RS-AS3VM-AL algorithms in three different learning strategies. After integration of spectral clustering, then sum up the SVM-AL-DC, SSAIR-DC, RS-AS3VM-AL-DC algorithm.(4) Based on the experiment content, we design a CBIR retrieval system and complete the entire experiment on it.By comparison with experiment results, we verify semi-supervised active learning methods are superior in the retrieval results. After integration of spectral clustering analysis into it, it plays better in retrieval performance. All in all it has positive meaning.