Research On Bidirectional (Top-down/Bottom-up) Fusion Visual Attention Model

Visual attention is an important visual mechanism of human visual system.Human can quickly and accurately pay attention to the target area in complex scene ignoring the interference that is because of the visual attention mechanism for adjusting and controlling.How to model and calculate the human visual attention through computer simulation has a great role in promoting the development in the field of machine vision.Research on visual attention model involves video and image processing,neurobiology,neuropsychology,bionics,computer science and other multi-disciplinary theory knowledge.According to the generation mechanism,visual attention model can be divided into the bottom-up and top-down model.Bottom-up visual attention is fully data-driven and produced by visual stimuli.Top-down attention is task driven and independent of the visual stimulus.At present,there are a lot of researches about bottom-up attention,and formed some classical attention model,but these models have a larger limitation of practical application because of without considering the top-down control.Top-down attention mechanism is too complicated to model.So scholars simply construct the top-down attention and combine with bottom-up attention and build bidirectional fusion attention model.Due to the defects of these models,they lead to poor fusion effect.According to shortcomings and deficiencies of above models,we proposed a Particle Filter based bidirectional fusion visual attention model on the basis of predecessors' research in this paper.By introducing the Particle Filter framework,our model is a scientifically rational bidirectional fusion model to fuse the top-down and bottom attention.Use the proposed model we put forward a task-oriented bidirectional fusion visual attention saliency estimation method.Test results show that the method is more efficient than existing models to estimate attention saliency.Besides we improved motion attention model and proposed a bidirectional fusion visual motion attention moving target detection method by using our model.Test on several global motion scene,the results show that the method improved the accuracy of target detection compared with other methods.