Visual Saliency Detection Based On Region Contrast

Digital image and video have been indispensable in our daily lives for informationtransmission and exchanging, along with the popularity of equipments with camera,such as digital cameras and smart phones. For computers, digital image and videoare just composed of random pixels, rather than identifiable objects and meaningfulscenes. Since digital image and video present various content, it’s hard for computersto understand the semantic information in image and video in the same way as humans.According to visual attention mechanism, visual saliency can detect the salient regionsin image and video and it plays the similar roles to distinguish and process the externalinformation as Human Visual System. With the help of visual saliency, computer candeal with regions on their saliency and even adaptively edit and process the image andvideo based on their salient content. However, Human Vision System is not yet fullyunderstood. Allowing computers to detect visual saliency on the basis of the knownvisual principles, is an important topic for computer vision and graphics.Obviously, different saliency detection algorithms would produce different salien-cy maps as a result of different views on visual disparity. And it deserves further studyto meet the demand of practical applications. This thesis focuses on visual saliencydetection based on region contrast, which is mainly in view of the bottom-up visual at-tention mechanism, and it can produce legible contour and shape of the salient regionsin the saliency map.The thesis firstly further discusses and summarizes the common principles ofvisual attention and usual features in the process of saliency detection. The visualfeatures include color, lightness, edge, motion, frequency, etc. Then in accordancewith practical demand, we analyze visual saliency detection algorithm based on re-gion contrast and find this framework is suitable to different visual features. Com-pared with image, the content of video is too complex. Also the video gives more cues to detect the salient region. On the basis of image saliency, we propose a videosaliency detection algorithm based on region contrast. This algorithm takes the advan-tages of inter-frame relevance to detect the object-like regions, and uses the dynamicfeatures,which consider color,motion and texture, and multi-scale region contrast toproduce the legible contour and shape of the salient region. This algorithm takes bothspatial and temporal factors into saliency detection and we has already evaluated thepracticability of our algorithm on the open test dataset.Based on visual saliency detection, we also apply the saliency to practical appli-cations. we divide the these applications into two categories:saliency-guided applica-tions and saliency-preserved applications. In this thesis, saliency-guided applicationscontain automation object segmentation, content-aware seamless composition, styliza-tion and special visual effects, and saliency-preserved applications contain decoloriza-tion and smart crop. Most saliency-based applications further prove the effectivenessof our saliency detection algorithm.