Research On Saliency Detection Based On Visual Intrinsic Attributes

Visual saliency is the result of human visual system selectively processing the important and noticeable visual information,and ignoring the inessential information under the direction of visual attention mechanism.In computer vision area,researchers have proposed many visual saliency detection approaches to detect the salient targets from complex scene and these methods have been used in many computer vision applications such as image segmentation,object recognition,and so on.Traditional saliency detection models ignore the influence of extrinsic environmental factors on image formation and fail to detect salient objects accurately when the background is similar to the foreground.Essentially,images are combination of several physical intrinsic attributes which can overcome the influence of environmental and enhance image discrimination.According to the demands of computer vision research for saliency detection,in this thesis,we research the key technique of visual saliency detection based on intrinsic decomposition theory and attributes.The main work in this thesis is listed as follows:(1)We introduce several widely used features in saliency detection and the related intrinsic attributes.Also,we introduce several classical detection models in terms of bio-inspired,computed-based and task-driven,then summarize their advantages and disadvantages.(2)In order to overcome the influence of extrinsic environmental factors and intrinsic similarity between foreground and background,we take image intrinsic attributes which are invariant to environmental factors into count.The variations of image appearance are exploited to categorize salient and non-salient pixels,and regional color contrast computed on multi-size superpixels is also computed to enhance the uniformity of salient regions.Finally,we construct an integrated framework by jointing pixel-level gradient-based and regional color-based saliency.The proposed method is simple,efficient and robust enough to estimate the foreground object against background that with similar appearances.Experimental results demonstrate that our method can achieve comparable performances with state-of-the-art methods.(3)In order to extract the foreground motion information and suppress the background in dynamic scene,we construct a video saliency detection approach based on spatio-temple saliency detection.We employ optical flow method to acquire motion information from video sequence and measure the temporal saliency by computing regional motion contrast within each frame.Furthermore,to overcome the influence of background motion information on video saliency detection and generate the final saliency maps,we construct an adaptive fusion method by analyzing the extent of object motion which dynamically weighted the temporal and spatial saliency map.Also we conduct experiments on public available video datasets to verify the feasibility of our method.