Researches On Visual Saliency Computation And Its Applications

With the rapid development of Internet, the amounts of images and videos are growingexplosively. This information brings new challenges to people’s lives as well as convenience.On one hand, limited computational resource is needed to be allocated to the target of interest.On the other hand, the results of visual computing should comply with people’s requirement.Visual attention is an important facet of human vision in everyday life. It makes it possible toprocess the complex visual scenes and to locate the target of interest. Inspired by humanvisual information processing, visual attention mechanisms are introduced into the process ofcomputer vision information processing, which can meet the above two requirements. In thisthesis, the key technologies of visual saliency calculation and its applications are studiedfrom the aspect of visual attention of human being. At the same time, some new ideas andalgorithms of visual saliency computation are proposed. The main work and contributions ofthis thesis are as follows:Firstly, this thesis proposes a bottom-up visual saliency calculation method based onimage sparse representation, which can be respectively applied to the saliency calculation ofstatic images and video frames. For the saliency computation of static image, the proposedmethod first gives a sparse representation of image. Then it computes the local saliencyaccording to the similarity of sparse feature between the central block and the surroundingblocks of the image. Next, it calculates the global saliency according to the similaritybetween the central block and the other blocks of the image. At last, saliency map isgenerated by merging the local saliency and the global saliency. Comparison experiment onthe public data sets shows that this method achieves better result on predicting the human eyegaze than other similar algorithms. For the calculation of saliency in the video, the proposedmethod considers both the saliency of images in the same frame and images in the differentframes as well as the merging of these two kinds of saliency. Experiment on the public videosshows that saliency map generated using this method can highlight significant goal in thevideo effectively.Secondly, this thesis proposes two kinds of saliency calculation methods consideringdifferent visual tasks. According to the visual search task in still images, this paper proposesa saliency calculation method based on vision gaze. By learning the relationship between the global features of invariant scale based on image sparse representation of scene and visiongaze, this method is able to model top-down saliency of search based on specific tasks.Considering the bottom-up factor, the method uses probability learning to manage theintegration of the top-down saliency and the bottom-up saliency. Comparison experiment onthe public data sets and self-built data shows that this method can achieve good gazeprediction effect. For target tracking in video, this paper proposes a saliency calculationmethod based on local color comparison. By learning the relation between the colorhistogram of target area and that of the surrounding areas, this method sets up the saliencymodel based on local color contrast, which can generate visual saliency map based on targetfrom new video frames. Experiment on the public videos shows that saliency figuregenerated using this method can predict target tracking gaze effectively.Thirdly, this thesis proposes a saliency target detection algorithm based on the globalcolor comparison. This algorithm introduces the global color contrast characteristics ofimages in the process of saliency map segmentation, and puts them in the conditional randomfield framework to mark salient objects, thus avoiding the insufficiency of saliency objectssegmentation only with saliency map. Comparison experiment on the public data sets showsthat this algorithm can segment the saliency targets effectively.Finally, this thesis proposes a target tracking based on visual saliency calculation. Thisalgorithm can determine the approximate location of the target by saliency calculation, whichhelps get the target search scope. Then it conducts multi-scale with target distribution fieldfeature in the search area, thus realizing the precise localization of target. This algorithm canadapt to the scale and the apparent of target change. In addition, this algorithm can also findocclusion, and can detect and position the target accurately when the target recurs.Experiment on the public videos shows that the proposed algorithm can cope with occlusionduring target tracking and change of scale effectively.