Supported by the National Natural Science Foundation of China under Grant Nos.60721002, 60373065, 69903006 (国家自然科学基金); the National High-Tech Research and Development Plan of China under Grant No.2007AA01Z334 (国家高技术研究发展计划(863)); the Program for New Century Excellent Talents in University of China under Grant No.NCET-04-0460 (新世纪优秀人才资助 计划)
This paper proposes an algorithm named Small Sample Incremental Biased Learning Algorithm to solve three difficulties of relevance feedback in sketch retrieval, including small sample issue, asymmetry of training data and real-time requirement. The algorithm combines active learning, biased classification and incremental learning to model the small sample biased learning problem in relevance feedback process. Active learning employs uncertainty sampling to choose the best labeling samples, so that the generalization ability of classifier is maximized with the limited training data; Biased classification constructs hyperspheres to treat positive and negative data differently, which distinguishes the user’s target class accurately; Newly labeled samples in each feedback loop are used to train the classifier incrementally to reduce the training time. Incremental learning also collects training data to further alleviate the small sample problem. Experimental results show that this algorithm improves the performance of sketch retrieval. And it can be well extended to other retrieval domains like CBIR (content based image retrieval), 3D retrieval, and so on.