Learning from user feedback for image retrieval

Relevance feedback technique has been one of the most active research areas in the field of content-based image retrieval. In this paper, we use Gaussian mixture model to represent the user's target distribution, which can further narrow down the gap between high-level semantic and low-level features. Furthermore, we present a novel approach to estimate the distribution parameters based on the expectation maximization algorithm. Because current image retrieval systems are incapable of capturing user's inconsistent intentions, we propose a framework to resolve user's conflict feedback. Experimental results show that our system can gradually improve its retrieval performance through accumulated user interactions.