Hierarchical correlation for content-based image retrieval

This report proposes a hierarchical correlation calculation approach to content-based mammogram retrieval. In this approach, images are represented as a Gaussian pyramid with several reduced-resolution levels. A global search is first conducted to identify the optimal matching position, where the correlation between the query image and the target images in the database is maximal. Local search is performed in the region comprising the four child pixels at a higher resolution level in order to locate the position with maximal correlation at greater resolution. Finally, this position with the maximal correlation found at the finest resolution level is used as the image similarity measure for retrieving images. Experimental results have shown that this approach achieves 59% in precision and 54% in recall when the threshold of correlation is ≥0.5.