Vision Attributes Based High Dynamic Range Image Processing

The range of light in real scenes,from sunlight noon to stary night,can be huge over 15 orders.the extraordinary human vision system can distinguish most of the details,while the traditional single channel low dynamic range images only have 256 levels,which is far smaller than the scenes.the emergence of the high dynamic range techniques makes it possible to store and display images with large luminance variations.HDR images are widely used in movies,games,simulations,virtual reality and military applications due to its high quality.However,because its dynamic range is far larger than the existing consumer displays,how to visualize HDR images is an important research direction.Tone mapping is a computer graphics technology that displays high dynamic range images on a limited dynamic range device.The main content of this thesis is the tone mapping algorithm based on visual mechanism.Most of the existing algorithms are based on the edge preserving filters,these methods can extract the base layer of the image quickly and effectively.However,due to the defects of the edge preserving filters,the results of tone mapping will produce halo artifacts,and make the details in extremely bright and dark areas not be significantly improved.In order to solve these problems,two effective and stable tone mapping algorithms are proposed.(1)Tone mapping algorithm based on local extrema.The traditional edge preserving filters always treat edges as positions that have drastic gradient changes.But,the existences of texture with big gradient changes and the edge with small gradient changes may lead to wrong smoothing results.In this thesis,details are defined as small oscillations among local extrema.This method can distinguish the details with high contrasts from the edges with low contrast,so that edges in the image are preserved and textures are smooth.Results of this tone mapping have no obvious halo artifacts.(2)Tone mapping algorithm based on side window filtering.Existing filters are based on full window regression,that is,the center of the window is placed on the pixel to be processed,but when the pixel is located on the edge,the filter will inevitably blur the edge.The side window filtering technology only considers the pixel on one side of the edge,avoiding the influence of the pixel on the other side,so as to achieve a good mapping effect.Experimental comparison and analyses show the two algorithms proposed can effectively restrain the halo artifact,and have obvious enhancement effect on the detail information of extremely bright or dark.