Learning from Relevance Feedback Sessions using a K-Nearest-Neighbor-Based Semantic Repository

This paper introduces a flexible learning approach for image retrieval with relevance feedback. A semantic repository is constructed offline by applying the k-nearest-neighbor-based relevance learning on both positive and negative session-term feedback. This repository semantically relates each database image to a set of training images chosen from all semantic categories. The query semantic feature vector can then be computed using the current feedback and the semantic values in the repository. The dot product measures the semantic similarity between the query and each database image. Our extensive experimental results show that the semantic repository (6% size and 1/3 filling rate) based approach alone offers average retrieval precision as high as 94% on the first iteration. Comprehensive comparisons with peer systems reveal that our system yields the highest retrieval accuracy. Furthermore, the proposed approach can be easily incorporated into peer systems to achieve substantial improvement in retrieval accuracy for all feedback steps.