Study On Selection Of Focus Of Attention And Optimal Trajectory Planning Of Gaze Shifts For Modeling Visual Attention Mechanisms

Human visual system can priority assign the limited computing and storage resources to a few regions of interest in the scence, which solves the“bottleneck problem”of visual information processing, and this specific property is called visual attention mechanism. If this critical mechanism can be applied to the fields of image analysis or computer vision, the efficiency of processing image and video information would be greatly improved. Thus a consierable amount of researches have focused on this subject and numerous computational models of visual attention have been developed. The major challenge for modeling visual attention mechanisms is to select focus of attention (FOA) and optimally plan trajectory of gaze shifts. At present extensive researches have been done on selecting the FOA, but there are few researches on trajectory planning of gaze shifts. Unfortunately, those two problems have not been solved very well. Thus this paper will take some new attempts in those two aspects, and the work we will do are as follows:First, we propose a method to generate a saliency map based on the detection of the keypoints, and determine the location of the attention focus combined with image segmentation. As the keypoints of the scale space mean that there is some important information in the local range of this keypoint, so we propose a method to generate a saliency map using the scale and location information of the keypoints. And the result shows that the local saliency of image is related to the visual perception. The difference from other methods is that our method increases the image segmentation. The greatest benefit of that process is that our method can effectively determine the number of attention focus, and maintain a certain distance between the attention focus, and avoid the repeated attention of the same object or region. Then we can find out more image information with less attention focus.Second, we propose a method to design the trajectory of gaze shifts based on graph theory. This method skillfully turns the problem of designing the trajectory of gaze shifts into Hamilton path optimization problem, so it makes the trajectory of gaze shifts better abided by the rule of attention focus location shift. In order to mediate the characteristics of descending and adjacent priority which are conflict with each other, we propose a new optimization index named the connection cost. And the result shows that this method is superior to“Winner-Take-All”neural network, which is a widely used method.