The Research On Key Technology Of Computational Model For Visual Attention

Visual attention mechanism is an intrinsic property of human and other primates. With the help of visual attention mechanism, human vision system can optionally process the visual information and the contradiction between the limited resources and the huge visual information can be solved effectively. By using visual attention mechanism in the research fields such as image processing, pattern recognition and machine vision, the amount of processed information and computational resources can be reduced and the efficiency of information processing can be increased effectively. Therefore the research of computational model for visual attention is the hot issue in these research areas. Based on the existing research results, computational model of visual attention and the key technologies are intensively studied in this paper. The researches of the dissertation are:(1) The calculation of visual saliency is one of the key problems of the computational model for visual attention and it is the basis of the selection and shift of attention. The computational method of visual saliency is studied in this paper to solve the problems of the existing approaches. A new approach is proposed in this paper. The proposed approach computes the visual saliency from three aspects which are local saliency, global saliency and rarity saliency. The proposed approach solves the problems of the existing methods effectively and the accuracy and effectiveness are increased significantly.(2) Visual saliency is the result of competition among multiple visual features. After calculating the conspicuous map of these features, it is required to integrate them to generate the integration saliency map. Therefore, the integration method is studied in this paper and an improved method of feature integration is proposed. According to the compactness and spatial distribution of salient regions, the weights of the feature conspicuous maps are computed dynamically. The problems such as fixed weights or getting weights through learning of the existing feature integration strategies are solved effectively.(3) The selection and shift of focus of attention is another key issue in computation model for visual attention. The main rules of the current methods are the proximity priority and inhibition of return, but they are not completely consistent with human vision system. A depth-first and hierarchical mechanism of focus of attention selection and shift is proposed in this paper. Firstly, the input image is analyzed with the smallest scale. When the first attention region is selected, it is processed as the input image using the same method and other regions are ignored. After the first attention region is completely analyzed, the focus of attention is shifted to the original image, the next attention region is selected according to the principle of proximity priority and inhibition of return and the process described above is repeated. The hierarchical selection and shift mechanism not only is in accordance with human vision system but also can simulate the covert attention and overt attention together.(4) To solve the difficulty of calculating the size and shape of attention region in space-based visual attention, the computational model for object-based visual attention is studied in this paper. The method of extraction of perceptual object based on spatial saliency is proposed. The calculation method of visual saliency based on perceptual object and selection and shift method of attention are also proposed. Comparing with space-based computational model, object-based attention model ensure the completeness of the attention object and the probability of shift to meaningless region is less. In addition, the hierarchical attention shift strategy is used in the paper and it is more consistent with human visual system.