Researches On Synergetic Computational Model For Visual Selective Attention

Visual selective attention computational model is a hot research issue in active vision. The research contents of visual attention are so abundant refering to visual physiology, psychology, computer vision, artificial intelligence and information processing. The research on visual selective attention mechanism provides the basis for designing more intelligent animate vsion system. The selected interesting region/object fastening the system resoureses to the salient regions, it facilitates the following visual processing such as image understanding, scene analysis, object detection and recognition.The sygergetics theory and sygergetic pattern recognition method is utilized to research the pivotal issues in visual attention computational model. The goal of this thesis is to build a synergetic object recognition model based on visual attention mechanism. The thesis is organized in the clue as "Representation ->Learning->Attention->Selection ->Recognition". Firstly, the non-linear scale space representation of visual feature images and optimal scale selection method of attention area is researched and built. Secondly, bias between visual features influenced by the task is calculated through the synergetic model perception method. The data and visual task is fused by the task-relevant salient map. Then, the speciality of perceptual object as the unit of visual attention is researched. The visual attention salient measurement based on object which integrates the space-based and object-based computational model is proposed. Finally, the synergetic object recognition model piloted by visual selective attention mechanism is built.The main work of this thesis includes:(1) Status of visual selective attention mechanism in both cognition and engineering are summarized. The problems existed in current visual attention computational model are pointed. The feasibility of researching visual selective attention mechsnism by synergetics and synergetic pattern recognition is analyzed. The research connotation and method of synergetic visual attention computational model are proposed.(2) The center-surround computational strategy of visual receptive field and integrated field is deeply analyzed. Based on the research of image scale space, non-linaer scale space representation of visual attention feature image and the optimal scale selection of attention area are proposed.(3) The perception of ambiguous and multivocal model is researched by synergetic recognition theory and synergetic model perception method is proposed. Bias between visual features influenced by the task is calculated through the synergetic model perception method. The salient regions selected by task-relevant salient map are more correlating to the task and the transformation of attention point is more reasonable. The task-relevant salient map based on sygergetic model perception combines the bottom-up and top-down visual attention research. (4) The definition, modelation and detection algorithm of perceptual object as unit of visual attention are researched. Acorrding to the basic ideology "Cooperation forms structure, competition promotes development" of synergeitcs, the grouping strategy of perceptual object is reaesrched. The object-relevant visual attention salient measurement method is finally proposed which relates the spcace-based and object-based visual attention research.(5) Contraposing the object detection and recognition in complex scene, an object recognition model based on visual attention mechanism is proposed. The prototype object representation based on PCA and synergetic learning algorithm is researched. In order to solve the invariance recognition of salient regions, a synergetic invariance algorithm based on influentuial dynamic is proposed. The proposed method integrates the visual attention mechsnism with object recognition and is very consistent to the cognition process of human in representation, lenaring, attention, selection and recognition.