Research Of Visual Co-saliency Object Detection Algorithm

With the rapid development of Internet, the popularity of smart phones, the social networking sites, the applications for image and video and other visual data is growing rapidly in the information age. Faced with such a flood of information,people tend to get more useful information from the image and video data via computer. The analysis of visual saliency can simulate the human image of visual perception, and detect the significant object or area which compliance with human visual perception so that people can allocation the limited computing resources to the important part of the visual information and analysis and processing these data, this will help shorten the data processing time of the system, and improve the information processing efficiency. So far, human cognition of their visual system is not enough,and there are many deficiencies of the existing visual salient model, the research of the visual salient model is still one of the hot spot in the computer visual field.In this thesis, the main research content includes the following two aspects一:(1) When given a set of images containing a specific object, the human visual attention will increasingly focus on such object. In order to simulate this visual perception of human effectively, this thesis proposes a visual salient object detection via weakly supervised learning. The detection algorithm can enhance the learning and updating of the salient region so that it can learn the appearance of specific object,then using it to detect the visual salient object. This algorithm can simulates human gradually attention and learning process for a specific object. The experiment results show that the object detection algorithm can detect the visual salient object and it turn a high accuracy.(2) Human’s perception of things can be divided into pre-attention stage and the recognition stage, in the case of the unconscious, the discriminate most region or object in the image can capture our attention, but after people have knowledge about the visual object, then our vision will pay more attention to the visual object related to prior knowledge. To simulate this human visual perception process, this thesis proposes a visual salient object detection algorithm based on the collaboration perception of the bottom and top information. The algorithm achieves the goal of visual perception by using the top-level information to change the distribution of the low-level visual salient features. The algorithm extracts a variety of primary visual features of the images firstly, and calculates the saliency map of a variety of primary visual features, then obtain information related to the visual object by modeling on the priori knowledge of top, and achieve the goal of collaborative awareness by combining the low-level salient features and top-level information. Experiment results show that the object detection algorithm can simulate the human visual perception process effectively.