This dissertation explores the concept of visual saliency---a measure of propensity for drawing visual attention---and presents various novel methods for utilization of visual saliency in video compression and transmission. Specifically, a computationally-efficient method for visual saliency estimation in digital images and videos is developed, which approximates one of the most well-known visual saliency models. In the context of video compression, a saliency-aware video coding method is proposed within a region-of-interest (ROI) video coding paradigm. The proposed video coding method attempts to reduce attention-grabbing coding artifacts and keep viewers' attention in areas where the quality is highest. The method allows visual saliency to increase in high quality parts of the frame, and allows saliency to reduce in non-ROI parts. Using this approach, the proposed method is able to achieve the same subjective quality as competing state-of-the-art methods at a lower bit rate. In the context of video transmission, a novel saliency-cognizant error concealment method is presented for ROI-based video streaming in which regions with higher visual saliency are protected more heavily than low saliency regions. In the proposed error concealment method, a low-saliency prior is added to the error concealment process as a regularization term, which serves two purposes. First, it provides additional side information for the decoder to identify the correct replacement blocks for concealment. Second, in the event that a perfectly matched block cannot be unambiguously identified, the low-saliency prior reduces viewers' visual attention on the loss-stricken regions, resulting in higher overall subjective quality. During the course of this research, an eye-tracking dataset for several standard video sequences was created and made publicly available. This dataset can be utilized to test saliency models for video and evaluate various perceptually-motivated algorithms for video processing and video quality assessment.
展开▼
