Research On Method Of Salient Region Detection

Computer Visual has attracted attention of researchers for a long time and relevant technologies have involved in our daily life. In the face of the huge image information in this real world, processing efficiency of machine is often far from satisfaction, while actually human eyes are only interested in a small area of the image. Therefore, if we can use the machine to mimic human eyes to identify these areas of interest, then we can increase the efficiency of visual processing. The salient region detection is to find a possible target area, narrowing the scope of the search and laying the foundation for subsequent work. The reliable saliency detection algorithm can provide valuable reference data to target detection, target tracking and recognition. In order to find suitable and effective salient region detection model, the paper focus on three aspects: measurement method, feature fusion and utilization of the prior information about the object.Since the mechanism of salient region detection is based on the model of visual saliency, in order to have a fundamental understanding of the related issues about salient region detection, this paper summarizes the knowledge of neurophysiology and cognitive psychology which are related to the visual saliency. It includes visual signal conduction process in the optic nerve system, the function of visual pathway in the visual signal processing and the cognitive psychology phenomenon of bottom-up and top-down saliency.In order to measure the salient region of image, we proposed a method of salient region detection based on "diffusion maps". This method extracts the features and creates the graph model based on the superpixels. Mapping the graph data in manifold space by diffusion maps, computing diffusion distance on manifold space to measure the difference between the superpixels and the global contrast information, thereby calculating the saliency. The experimental results show that the saliency calculating on the manifold have a good measure of the differences between the image regions, effectively highlights the region of saliency.According to the feature fusion of salient region detection, we proposed a method of features. Due to the fact that expression on the region is different when the feature is different, so saliency value based on it is also different. If we can integrate these features in the reasonable way and make them complement each other on saliency expression, then we can improve the quality of saliency. The "cross diffusion" algorithm makes the similarity matrix with the different feature fuse with each other and normalize the result of fusion to get the transition matrix work on the initial saliency. Experimental results show that this method of feature fusion can make good progress on the saliency value in most cases.The top-down salient region detection model depends on the prior information about object. We proposed two methods of priori information about object for the saliency computation. Firstly, we use the label information about the training sample to acquire the improved distance metric through the supervised metric learning algorithm. Then we get the enhanced saliency value by global contrast. Secondly, we use the objectness detection method to get the candidate windows of objectness in the picture. By the coordinates of these windows we acquire the location probability distribution of object. Choosing this location probability distribution to be a weight to correct the initial saliency can make a more precise saliency value. In the result, the salient region in the saliency map is more outstanding, the outline is clearer, and the inhibition of background is better.