CINDI at ImageCLEF 2006: Image Retrieval & Annotation Tasks for the General Photographic and Medical Image Collections

This paper presents our techniques used and their analysis for the runs made and the results submitted by the CINDI group for the task of the image retrieval and automatic annotation of ImageCLEF 2006. For the ah-hoc image retrieval from both the photographic and medical image collections, we have experimented with cross-modal (image and text) interaction and integration approaches based on the relevance feedback in the form of textual query expansion and visual query point movement with adaptive similarity matching functions. Experimental results show that our approaches performed well compared to initial visual or textual only retrieval without any user interactions or feedbacks. We are ranked first and second and achieved the highest MAP score (0.3850) for the ad-hoc retrieval in the photographic collection (IAPR) among all the submissions. For the automatic annotation tasks for both the medical (IRMA) and object collections (LTU), we have experimented with a classifier combination approach, where several probabilistic multi-class SVM classifiers with features at different levels as inputs are fused with several combination rules to predict the final probability score of each category as image annotation. Analysis of the results of the different runs we made for both the image retrieval and annotation tasks are reported in this paper.