Researches On Visual Saliency Detection

Visual saliency detection is a process to extract regions or objects that are attracted to the observer in the scene based on the visual attention mechanism. It is one of the important problems in the field of image analysis and computer vision. Visual saliency detection can be widely used in the field of image segmentation, image retrieval, image coding, and object recognition. The research of 2D visual saliency detection is growing rapidly at present, while the research of 3D visual saliency detection is just unfolding with the development of stereo technology. This thesis researches deeply on visual saliency detection under the above background.This thesis first summarizes various methods of 2D visual saliency detection. The theory and technology of 2D visual saliency detection are studied from the aspects of concept, classification, construction of models, and evaluation method. On this basis, this thesis proposes a universal model for saliency detection. The saliency map of any existing saliency detection methods is utilized as a priori knowledge for the universal model. The Bayesian decision theory is adopted to refine the rough saliency map to obtain a more accurate saliency map. Meanwhile, an iterative optimization strategy is designed to obtain better saliency results. The universal model can effectively improve the performance of saliency detection.In the aspect of 3D visual saliency detection, this thesis realizes a superpixel and regional contrast-based 2D visual saliency detection methods, and proposes a 3D visual saliency detection model through introducing the depth information. The experimental results show that the proposed 3D visual saliency detection method has good performance, and can accurately extract the salient objects.