Query movement by sensitivity of similarity measure

In Content Based Image Retrieval (CBIR) system, Relevance Feedback (RF) technique can substantially promote retrieval accuracy. As a popular algorithm of Relevance Feedback, Query movement was widely applied in this area and it moves the query in the input space towards relevant images. However, traditional Query Movement algorithm has significant disadvantages. It is to reduce the similarity between the query image and irrelevant images and increase similarity between the query image and relevant images. However, the magnitudes of increment or decrement of similarity differ for different images. Some sensitive images in the database may yield more increment or decrement of the similarity of the query image. In this work, we propose a sample sensitivity based query movement. Sensitive images in the database will contribute more to the movement of query. The sensitivity of image is measured by a stochastic sensitivity of the similarity measure for abstract unseen images located within a Q-neighborhood of an image. Experimental results show that the proposed method outperforms simple query movement without sensitivity.