Improving image retrieval effectiveness via sparse discriminant analysis.

The semantic gap between low-level features and high-level semantic concepts is a fundamental problem in content-based image retrieval (CBIR). To close this gap, relevant feedback is included in the CBIR. Based on the user's feedback samples, a projection matrix is learned to project samples from the multi-dimensional original space to the low-dimensional projection space, and a classifier is learned on the projection space to classify the images. However, the number of classes in the relevance feedback is very small (only two classes), which leads to low classification performance and results in poor retrieval performance. To solve this problem, we propose a novel supervised image retrieval method, called Sparse Discriminant Analysis for Image Retrieval (SDAIR). Different from existing image retrieval methods, which have low precision due to small-class size problems, SDAIR is designed to not be affected by small-class size problems. Therefore, SDAIR is potentially more suitable for image retrieval with relevant feedback, where class sizes are often very small. The experimental results on the two databases demonstrate that the proposed method obtains competitive precision compared with other content-based image retrieval methods. [ABSTRACT FROM AUTHOR]