Visual Saliency Detection And Its Application In Video Coding

Visual saliency detection aims to predict relatively outstanding regions, which draw human visual attention, in an image or video. Saliency modeling simulates the human visual system by building data simulation models with a variety of video or image characteristics. This thesis mainly studies human visual attention mechanism and characteristic in the frequency domain of image to predict regions of interest.The thesis first analyzes advantages and disadvantages of previous saliency detection models from domestic and overseas, and then proposes two novel bottom-up saliency detection models based on the prior knowledge of the human visual attention mechanism.The fist proposed saliency model is an improved saliency detection method based on phase spectrum of Quaternion Fourier Transform. The human visual attention weight matrix is applied to adjust the portion of the image quaternion characteristics consisting of intensity, color and motion features. The phase spectrum of the Quaternion Fourier transform is used to obtain the saliency map of an image.The second proposed one is based on the human visual sensitivity and spatially weighted dissimilarity of discrete cosine transform(DCT) coefficients. To compute image contrast, we propose to generate a feature vector from Discrete Cosine Transformation coefficients for patch representation. The saliency score for each image patch is calculated by not only the dissimilarities between the DCT coefficient of target patch and other patches in the whole image, but also the visual impacts for these dissimilarities weighted by the human visual sensitivity.Together with the foveated just-noticeable-distortion(FJND) modeling, our proposed saliency models are applied to ROI-based H.264 and HEVC(High Efficiency Video Coding) video coding. The experimental results demonstrate that the proposed saliency models consistently outperform the state-of-the-art models in three widely used eye-tracking benchmark datasets. In addition, the proposed saliency-based video coding algorithm also achieves better performance than conventional coding methods, especially for those video sequences containing obvious salient objects.