Bayesian Multi-instance Learning for Image Retrieval with Unlabeled Data
Tao Chen, Huifang Deng · Proceedings of the 2nd International Conference on Computer Science and Electronics Engineering (ICCSEE 2013) · 2013
To deal with the two problems in image retrieval, i.e., the small number of query images, the ambiguity of an imagethe image consists of many regions with different semantic meaning, in this paper, we proposed a novel method for image retrieval based on Bayesian multi-instance learning using unlabeled data, termed as Bayesian-MIL method, which treats the image retrieval as a binary classification problem.In this method, to obtain an approximate estimation of the classconditional probability of positive images, a multi-instance learning algorithm is adopted to filter out background regions in positive images, and then a Bayesian classifier is constructed to rank the images from a large digital repository according to their score of posterior probability.Finally, the ranking top k images will be returned to users.Experimental results on COREL image data set have demonstrated the effectiveness and efficiency of the proposed approach.