Evaluation of Global Descriptors for Multimedia Retrieval in Medical Applications

In this paper, global descriptors from MPEG7, GIST and Compact Composite Descriptors are evaluated for image retrieval in the IRMA-2007 medical collection. This evaluation tests descriptors using every image from each class instead of a small group of representative images. The evaluation results obtained by Mean-Average Precision (MAP) and precision@N indicate that MPEG7 EH, GIST and Fuzzy BTDH outperform the other global descriptors analyzed by a large margin, even more when combined by late-fusion rank aggregation. A multimedia retrieval evaluation system was developed to support the experiment and offers the possibility of textual, visual and combined searches over the medical collection.