Research On Technology Of Salient Object Detection In The Visual Filed

Inspired by human vision system can find the most interesting region in cluttered scene quickly and accurately, the technology of salient object detection in visual filed detects salient target automatically through imitating human visual attention mechanism. It has important theoretical meaning and wide application values among target auto-detection and recognition in the field of computer vision. We research on visual salient object detection in natural scene to investigate the fundamental principles and mechanisms in visual saliency detection, and build corresponding computational model. For two technical problem among current visual salient object detection algorithms namely real-time and accuracy, we are devoted to improving the accuracy. We adopt bottom-up technical route and create proper computational model to detect salient object in images and videos.The paper firstly introduces the typical image graph models which are used for detecting image salient object in image. On the basis of others, we propose an improved algorithm combined with absorbing Markov chain and manifold ranking. Difference from the previous graph models, we consider the characteristics of the target and background and introduce the concept of the absorbed time in absorbing Markov chain to detect salient object. Further,considering this algorithm cannot deals with the scene where background occupies larger area in the center of an image well. So we use manifold ranking to suppress background, then improve the precision. Also in order to improve running speed, we firstly use SLIC algorithm to segment an image into some superpixels and regard superpixels as the basic elements for subsequent processing; then define a score function to automatically determine whether to further detect salient objects based on manifold ranking. The algorithm not only ensures the real-time but also improves the accuracy.Next, we detect salient object in videos. For only utilizing color feature can not deal with some scenes, the paper considers the characteristics of videos, then introduce two new visual feature, namely motion feature and object probability distribution. We first compute three saliency maps corresponding to three features and then set weight coefficients adaptively with its entropy to fusion three saliency maps. The algorithm can detect salient object in many scenes well.The paper uses standard dataset corresponding to two algorithm for test and compare it with some classical mainstream visual salient object algorithms. Experimental results demonstrated that the proposed algorithms have higher accuracy. But these algorithms must depend on thehypothesis that target must have salient feature relatively which restrict its detection capability,we need develop further research in this filed afterwards.