Study on forensic image retrieval

Abstract

This paper compares the performance of different image features and that of different similarity measurements in a content-based image retrieval (CBIR) system using forensic images as the testing gallery. This special gallery contains 400 forensic images in 8 categories. In the experiment, color, texture and color-texture features of the images were extracted and compared. And with these feature vectors, different similarity measures were used to evaluate the similarity level between the query image and the image in the gallery. And the retrieval results on real-world Crime Scene Investigation images, evaluated by precision and recall, lead to a conclusion that city block distance as the similarity measure shows a better performance than the Euclidean distance that we usually use. In comparing with the standard database, Corel, we make an analysis of the results, which shows the specialty of forensic images.