Research On The Computation Models Of Attention Mechanism In The Visual Information Processing

Visual attention mechanism is one of the important problems in computer vision community, and it almost includes perceive science, nerve science, biology and all the subject of computer science. The process of visual attention mechanism is so complex that people don't understand all the process until now. Most of current visual attention models are data-driven bottom-up models. Although those models are successful in some ways, there are some shortages in many applications. For instance, the attention is often affected by prior knowledge which is a top-down visual attention model; people's attention is more easily directed to a motive stimulus in a static scene, which is dynamic and static visual attention; saliency region is so important that it should adopt higher resolution in image compression, which is variable resolution image compression. Therefore, it is important to research for those computation models of visual attention mechanism.In the model of target and background contrast, we fuse all training targets into a target class and fuse all training backgrounds into a background class. Weight vector is computed as the ratio of the mean target class saliency and the mean background class saliency for each feature; for an attended scene, all feature maps are combined into a top-down saliency map with the weight vector by a hierarchy method. Then, the top-down and bottom-up saliency map are fused into a global saliency map which guides the visual attention.In the model of target itself character, low-level visual object's features such as color, intensity, orientation and texture are used and each feature is divided into some different parties (e.g., red, green and blue for color feature) in the training phase. All the features are extracted from object itself and do not depend on the background information. These features are represented by mean and standard deviation stored in long-term memory. In the attention phase, corresponding features are extracted in the attended image. For each feature, the similarity map is obtained by comparing training feature map and attended feature map. The more similarly, the stronger of the similarity map. Then all the similarity maps are combined into a top-down saliency map. In the same time, a bottom-up saliency map is acquired by the contrast of attended image itself. Then, the top-down and bottom-up saliency map are fused into a global saliency map.In the model of dynamic and static salience, we propose a spatiotemporal saliency attention model based on entropy value. The input video is divided into some continuous frames. For each frame, low-level visual features such as color contrast, intensity contrast, orientation and texture are used. The entropy value map is obtained by calculating the entropy value of each point. All the entropy maps are normalized and are fused into a dynamic saliency map. The static saliency map is acquired according to bottom-up method. Then, the dynamic and static saliency map are fused into a global saliency map.In the model of variable resolution image compression, we use bottom-up model to obtain salience regions. Original resolution is retained in the first salient region; the lowest resolution is applied in the unapparent salient regions and the middle resolution is decided by the saliency order from high to low. By this method, we achieve variable resolution image compression by the model of visual attention. The model of image compression not only can achieve high compression ratio in total image but also can keep high resolution in salient regions.At last, we summarize the presented work. According to the imperfect aspects, we analyze and discuss the future work.