Research Of Visual Attention Model For Complex Object

Selective attention is a vital mechanism employed by human cognitive system to select important or interesting information from a large amount of perception information. It is also the important process of allocating processing resources of human brain. In the research field of computer vision, through simulating the visual attention mechanism of human visual system computer can finish complicated intelligent tasks which require extracting interesting information from complex scene image.The human visual system shifts attention according to the saliency of visual stimuli while scanning a scene. And the saliency of visual stimuli is not only determined by the stimuli themselves, but also influenced by the human brain. These two kinds of influences are called bottom-up and top-down factors in the research of visual attention mechanism. Accordingly, models which are stimulating the bottom-up attention mechanism are called bottom-up or data-driven models, while those which are modeling the top-down influence to attention mechanism are called top-down or task-driven models. In the thesis, we investigate some influential models of these two kinds of visual attention models.According to different theories and hypotheses related to human visual attention mechanism in cognitive neuroscience research, many research branches and directions in top-down visual attention modeling emerged. However, currently most of this kind of models have their limitations, and are not quite consistent with the behavioral performances of biological mechanism. In the thesis, we focus on how to integrate comprehensive target information influences into top-down visual attention model, and present a new visual attention model based on object structural representation, which employs salient blob-graph to represent target and guides attention to shift to areas where target probably exists. In the new model, we propose an encoding method for blob characteristic code and a calculating method for its similarity measurement based on rank correlation theories, adopt a kind of simplified adjacency list to represent the graph structure consisting of blobs, and design a blob search-and-merge strategy to fuzzily match blob-graph in order to output focal areas where target might be. Results of experiments of applying the model to visual search task show that the new model could effectively introduce multi-dimensional features and structure information of complex target into visual attention model, cover the given target more completely, reduce invalid attention shifts and thus make the model locate complex target faster. The proposed model could be used to locate complex structural target in natural scene images.