An effective noise-resilient long-term semantic learning approach to content-based image retrieval

This paper proposes a noise-resilient long-term semantic learning method for relevance feedback-based image retrieval. Our system accommodates erroneous feedback resulting from the inherent subjectivity of judging relevance, user laziness, or maliciousness. It also addresses three main drawbacks of traditional relevance feedback techniques. Specifically, it uses a statistical memory learning method based on the user's feedback to extract additional high-level semantic information between query and database images. The learned semantic relationship automatically adds potential positive images to the feedback set to improve SVM-based low-level feature learning. These two measures are seamlessly combined to compute the overall similarity between query and database images. Our experimental results on a 6000-image Corel database demonstrate the effectiveness of the proposed system.