Research On Visual Saliency And Object Recognition

In computer vision, using visual saliency model to detect saliency map has drawn significant attention from researchers. However, due to the complexity of foreground and background, it is a great challenge to accurately detect the salient region of an image. In recent years, the development of saliency theory makes it possible to select the ”useful” information for further calculation when detecting a target object in an image. It does not only reduce the computational complexity and memory overhead,but also increases the detection precision due to the exclusion of redundant information.Consequently, the study of how to apply the filtered data to object detection becomes very meaningful.We first proposes a saliency model based on background learning. The model follows two successive steps to detect saliency. The first step is to calculate the weight for each superpixel while the second step is to calculate saliency value for each superpixel by weighted local contrast. For these two steps, we first present a superpixel weighting model, including superpixel detection, feature extraction and weight prediction. Then we propose a weighted local contrast algorithm. The computational element in this algorithm is superpixel and the saliency value is represented by the color and spatial differences between each superpixel and all other superpixels. Experimental results show that our model can effectively improve the detection accuracy. Compared to other models, the saliency map detected by our model on MSRA1000 and ECSSD has higher resolution and average precision, the salient points distribute more uniformly in target area, and also, the border between foreground and background is much more distinct.Secondly, on the basis of other saliency based object detection algorithm, we present a novel object detection algorithm based on parameter learning. The algorithmfirst detects the saliency map for an image by our saliency model. Then it constructs a quality function on the saliency maps. The parameter of the function is given by solving a optimization problem, which is modeled by Bayes model. Finally, we use dynamic programming algorithm to calculate the extreme point for the function to locate the object’s position. Compared to the saliency based object detection algorithm proposed by [1], our algorithm has higher precision on MSRA1000 and ECSSD. Moreover, the detection results also show that our saliency model performs better than the saliency model proposed by [2] in saliency based object detection.