Finding regions of interest based on scale-space keypoint detection

One of the major challenges for modeling visual attention mechanisms is to extract visual cues for automatic detection perceptually important regions in a scene. Here, we propose a simple model for detecting regions of interest (ROI) inspired from keypoint analysis. We adopted the idea that the appearance of an interest region can be well characterized by the distribution of its local features (e.g., keypoints). ROI detection involves five steps: the input image is first decomposed into a set of low-level vision feature maps (e.g., intensity map, and two double-opponent color maps). Extrema in difference of Gaussian (DoG) scale space are then calculated for detecting the keypoints within each feature maps. The location and scale information of keypoints are integrated to create three conspicuity maps. These conspicuity maps are normalized and summed into an overall saliency map. Finally, a "small" number of salient locations are successively selected using a dynamical neural network. Experimental results show that the proposed model outperforms the Itti's model, a state-of-the-art competitive approach. © 2011 Springer-Verlag.