Robson Barcellos, Rogério Saranz Oliani, Luciana Tarlá Lorenzi, Adilson Gonzaga
Abstract. Color autocorrelograms have been shown to excel color histograms, color coherence vectors and color co-occurrence matrices when used as feature vectors in content based image retrieval (CBIR) systems. This is due mainly to their skill to detect the spatial relation of colors. In this work we show that further improvement in the performance of autocorrelograms can be achieved by choosing an appropriate color space. Specifically, when robustness to illumination condition changes is an issue, HSV color space has been proven to be a good choice to work with. Performance variations due to distance metrics and image data base size are also considered.
Figure 1. Precision X recall for 100 images and L2
Figure 2. Precision X recall for 100 images and L1
Figure 3. Precision X recall for 248 images and L2
Figure 4. Precision X recall for 248 images and L1