Research On Visual Attention Model With Joint Spatial And Feature Domain Information

The mechanism of attention selection is one of the important characteristics of human visual perception. If this mechanism in human vision system is introduced into the computer vision system, the performance and efficiency of computer-based image analysis is undoubtedly enhanced. To probe the method of computationally modeling the mechanism of visual attention and its application, the thesis focuses on the following aspects:First, the status of research on bottom-up visual attention model, top-down visual attention model, and the mechanism of fixation shift are reviewed. The attention model based application on image segmentation, image classification and object detection are also introduced. Three key factors as follows are drawn from further analysis on the mechanism of visual attention: dissimilarity between image patches, spatial distance between patches, and distance from each patch to the center of the image, namely central bias.Then, a visual attention model with joint spatial and domain information is proposed based on above three factors. The key step of the model is to weight the dissimilarity between two patches by spatial distance and central bias. Actually, the core idea is to define the strength of visual attention for each image location as rarity of global feature response weighted by spatial distance. Tests on several image datasets and video clips by using different measures demonstrate that fixations predicted by the proposed model are more consistent with human fixations than some other state-of-the-art attention models.Finally, the proposed model is applied to contend-based image retargeting. Although themes of images are often people, natural scenes and urban scenes, such high-level semantic information is unable to be detected by above model. Therefore detectors of human faces, pedestrians and vehicles are added into the model. The results show that the method of image retargeting based on the proposed model our-performs the method based on gray images on reducing distortion of foreground objects in images.