A Visual Attention Model Based On Bayesian Inference Using Multiple Cues And It’s Object Detection In Remote Sensing Image

As the rapid development of space detection technology and sensor technology, the spatial resolution of remote sensing image gradually improve, data quantity is also more and more big. Due to the limitation of hardware processing conditions, the existing data processing ability cann’t match the mass remote sensing image data. So how to find a small region of interest from the mass data, using the limited processing capacity for priority processing, in order to achieve a reasonable allocation of resources is a key problem in the field of remote sensing.Every time when open the eyes, the brain will receive a lot of outside scenery image information. However the brain can’t parallel processing all the information as its processing ability is limited. So the visual system uses visual selective attention mechanism to rapidly select a small amount and important visual information from a large number of visual information for priority processing, in order to reduce computational complexity. So it can be seen that the brain’s visual attention mechanism is important to remote sensing image processing.Most of the visual attention models are the improvement based on Itti classical model. However, they have a common problem that the saliency map is a linear superposition for feature maps using some kind of methods. Obviously, this way does not conform to the biological mechanism. Neuroscience research results show that the process of visual perception is a kind of bayesian inference process in brain. So how to construct a bayesian model accordding to biological mechanism becomes the new direction of visual attention. Chikkerur proposed a bayesian integrated framework in2010, using inference to combine the top-down and bottom-up information. The framework successfully simulates a variety of visual attention phenomenon. However its shortcoming is treating all the features the same, for example, in the framework,"red" and "yellow" has the same relationship with "red" and "square", but in fact the two relations must be different.This paper proposes a visual attention model based on bayesian inference using multiple cues, the model is based on the Chikkerur model, introduces the concept of visual cues, and combines with the new neuroscience research. The model is mainly composed of two parts which are visual cues extraction and bayesian inference integrated top-down and bottom-up information. Experiments focused on three visual cues: shape, color and context, and then allocation visual cues according to visual pathways. The shape and color cues which are related with object, occurred in the ventral path, and output bottom-up feature map. Context cue is related with space position, occurred in the dorsal pathway, provides the prior information of space position. The output of cues in the first part is the input of the second part. The first part imports in the bayesian network, so the model converts the acquirement of saliency map to the solvement of posterior probability. The experiment of object detection and location in remote sensing image using the proposed model and Itti model shows the proposed model can effectively detect and locate the object, has a better elimination the interference of background,and the generated saliency map is also better.