Saliency Computing Based On Depth And Texture Information

The rapid growth of image data has brought huge pressure to image storage,transmission and processing. Saliency computing aims at extracting regions of interestand allows preferential allocation of computational resource to salient regions insubsequent image analysis and synthesis to speed up processing and reducecomputing pressure. Extracted saliency maps are widely used in image segmentation,object recognition, adaptive compression and so on. Traditional image-based saliencycomputing methods only take texture information into consideration while ignoring3D geometry information of the scene, resulting in obvious error between thegenerated saliency maps and the actual observation.This thesis introduces depth information to help texture information in theresearch of image-based saliency computing methods. The main results are asfollows:(1) A saliency computing method based on imitated DOF is proposed. We firstbuild color histogram to contrast color, and then imitate the focal plane and the DOFunder the guidance of depth information to compute the saliency value of each pixel.The generated saliency maps are more layered and better match the characteristics ofhuman vision.(2) A saliency computing method based on region contrast is proposed. We firstgenerate and merge superpixels to segment images into regions, and then contrastdepth and texture information as well as spatial geometry information to compute thesaliency value of each region. The generated saliency maps not only ensure saliencysmoothness inside and contour completeness outside the objects, but also suppress thesaliency value of non-salient regions.Compared with the direct compression, the HEVC reconstructed results under theguidance of the generated saliency maps better preserve the clarity of details in salientregions and cost less encoding bits.