Semantic kernel updating for content-based image retrieval

A lot of relevance feedback methods have been proposed to deal with content-based image retrieval (CBIR) problems. Their goal is to interactively learn the semantic queries that users have in mind. Interaction is used to fill the gap between the semantic meaning and the low-level image representations. The purpose of this article is to analyze how to merge all the semantic information that users provided to the system during past retrieval sessions. We propose an approach to exploit the knowledge provided by user interaction based on binary annotations (relevant or irrelevant images). Such semantic annotations may be integrated in the similarity matrix of the database images. This similarity matrix is analyzed in the kernel matrix framework. In this context, a kernel adaptation method is proposed, but taking care of preserving the properties of kernels. Using this approach, a semantic kernel is incrementally learnt. To deal with practical constraint implementations, an eigendecomposition of the whole matrix is considered, and a efficient scheme is proposed to compute a low-rank approximated kernel matrix. It allows a strict control of the required memory space and of the algorithm complexity, which is linear to the database size. Experiments have been carried out on a large generalist database in order to validate the approach.