Saliency Detection Algorithm Research Based On The Visual Congext Feature Of The Task

With the advent of the information age，the demand and apply of the information andscience technology has becoming increasingly higher. Much Image and video informationto the the artificial data processing and bandwidth requirements has caused a lot ofdifficulties and pressure. According to the human visual attention system, while viewing apicture or a period of video, people often only concerned about the most interesting part.Therefore, we only deal with the region of interest of the image or with some kind ofpurpose or task to search for certain areas of the image. Thence, we only deal with theregion of interest of the image, it not only can meet the needs of human beings, but alsogreatly improve the efficiency of processing data. Therefore, the paper focuses on thetask-based visual attention mechanisms to carry out some of the new algorithm.Firstly, a global saliency visual attention model based on the color distribution of theprior target is proposed. First of all, the input image is segmented according to the contrastratio with the color distribution of different classes and the target to obtain the priorknowledge map constrained by target information. According to the color salient featureof different classes and the region contrast ratios of space distribution, the global saliencyis computed. The saliency map of the input image is obtained by combining priorknowledge map and global saliency map. Experimental results show that the proposedalgorithm is more reasonable.Secondly, the visual attention model of searching people which is combined withglobal sparsity and visual context is proposed. During the lower cognitive, computing thesparsity of the image to gain the global saliency. During the higher cognitive, extractingthe color descriptor, texton descriptor and the descriptor of target feature, which areguided by the visual context information, to gain the location of the target. The saliencymap of the input image is obtained by combining global salient map and visual context,and then the pedestrian region is extracted. Experimental results show that the proposedalgorithm is effective and reasonable.And finally, the saliency visual attention model of pedestrian which is combined with top-down and bottom-up is proposed. During the lower cognitive, computing the sparsityof the image to gain the global saliency map. During the higher cognitive, extracting thegist feature and the histogram of oriented gradient feature to gain the gist saliency mapand the target location map. During the phase of training, training all the features by liblinear svm to gain the best weight coefficient. Finally, the saliency map of the input imageis obtained by combining the feature and the weight coefficient, and then the pedestriansalient area is extracted. The experimental results show that the model can obtain moreaccurate detection region.