Mammogram Retrieval: Image Selection Strategy of Relevance Feedback for Locating Similar Lesions

2. Prior Work

To incorporate relevance feedback into content-based image retrieval, two main approaches are developed: the query point movement approach and re-weighting approach. The concept behind the first approach is to modify the query that is originally submitted by the user. It is assumed that there exists at least one image which completely conveys the intentions of the user, and its high-level concept has been modeled in low-level feature space (Kushki, Androutsos, Plataniotis, & Venetsanopoulos, 2004). The query point movement approach is to move the point of the query toward the region of the feature space that contains the ideal image (Zhong, Hongjiang, Li, & Shaoping, 2003). Based on this concept, the classic Rocchio algorithm was originally developed to improve the effectiveness of information retrieval system (Rocchio, 1971). The MARS system has applied the query point movement approach as one of methods for relevance feedback (Rui, Huang, & Mehrotra, 1998). The method used in the MARS system is called (term frequency-inverse document frequency), which generates pseudo-document vectors from image feature vectors and then applies the Rocchio algorithm to find the ideal point.