Research On Key Techniques Of Visual Saliency Detection

With the rapid developments of internet and multimedia technologies, the amount of digital multimedia is growing in a geometric way. Although the massive multimedia data bring great convenience in daily amusement, education and commerce, they also lead to many new challenges on the existing multimedia processing technologies. On one hand, the amount of multimedia greatly exceeds the processing capability of computer, people should allocate the limited computational resource to the important visual information. On the other hand, from the huge visual information inputted into human eyes, human can effortlessly selecte a small amount of important information for the further complex processing, people also hope computer can simulate this capacity. Saliency detection methods can automatically predict, locate, mine the important visual information, thus it can help computer effectively select important information from the massive visual data. Taking the requirements of computer vision tasks for visual saliency as the point of de-parture, this thesis focuses on addressing the key techniques of visual saliency detection. The main contributions of this thesis are described as follows:Firstly, this thesis presents a integrated feature based saliency detection method for predicting eye fixations. This method uses the way of generating integrated features to simulate the process of enormous neural cells being simultaneously tuned to different fea-tures, and it uses the saliency measurements of integrated features to obtain the saliency aroused by the output of the cells simultaneously tuned to different features. For more comprehensive saliency measurements, this method combines the local and global salien-cy measurements during the saliency computation process. The experimental results on the public eye fixations datasets show that this approach obtains good performance on predicting eye fixations.Secondly, this thesis proposes a background contrast based salient region detection method. This method analyzes and validates, regions without eye xations are very likely the image background. Based on the verification, this approach regards the com-plementary area of the convex-hull of eye xations as the possible image background, then it measures the saliency of each region by computing its contrast to the estimated image background. The experimental results on the public image dataset show that this approach can well highlight the whole salient region on the saliency map. Thirdly, this thesis presents a saliency density and edge response based salient ob-ject detection method. This method simultaneously considers saliency density and edge response the two properties owned by the salient object during the detection process. It formulates salient object discovery as the optimal window search to find the window with maximum saliency density and edge response scores, and employs the proposed saliency density and edge response based branch-and-bound search algorithm to locate the optimal window containing salient object. Finally, for extracting the salient object with a well-defined boundary, this method applies GrabCut method, and uses the located window to initialize GrabCut. Extensive experiments demonstrate that this method achieves well salient object detection results.Fourthly, this thesis proposes a foreground correspondence based co-saliency detec-tion method on multiple images. This method tries to locate the correspondent foreground regions between the multiple images, and generate a map highlighting the corresponden-t foreground region for each image. Then this method utilizes the background contrast based salient region detection method proposed by this thesis to generate the single-view saliency map for each image. Finally this method linearly combines the map spotlighting the correspondent foreground regions and the single-view saliency map to generate the co-saliency map. Extensive experiments on several datasets show that this approach can obtain good performance in various scenes.