Research On New View Synthesis Algorithm For Light Field Images

Light field images need a trade-off between angular and spatial resolution,which lower its practicability and require algorithms to enhance angular and spatial resolution separately.New view synthesis algorithm can be used to enhance the angular resolution of light field images,but existed methods focus on view interpolation,which makes the range of view synthesis very small and can not reconstruct the whole light field images under specified virtual camera pose.Meanwhile,when synthesize new view image outside of the exsisted light field images,there's no good solution to fill in the image black hole area produced by occlusion change.As for the problems,we conduct a study on the new view synthesis algorithm for light field images,main work includes:1?We analyze the difference between light field camera imaging system and traditional camera,and propose a total-variation based new view synthesis model for light field images,which generates new view light field images under specified virtual camera pose via back-warp and reprojection of subaperture view images.Occlusion is estimated by depth of the scene in the projection process,and meanwhile,pose change of the micro-lens linked with subaperture view image is taken into account when the light field camera pose changes by the model,thus can generate accurate projection results.2?As for the image black hole area produced by the occlusion change during view synthesis,we propose a light field image inpainting algorithm which combines disparity priority and image content distribution between subaperture view images.Image inpainting priority is computed by disparity of the image,and after the centural view image is inpainted,all supaperture view images are inpainted by the image content distribution between them,thus the image content consistency between subaperture view images is promised.3?We propose a deep convolutional neural network based view synthesis algorithm for light field images.Input light field images are first projected to desired viewpoint,and subaperture view images and masks in the same row are stacked to be the input of the network to achieve end-to-end light field image inpainting and optimizing task,which generated new view image that looks real and lower the time complexity of new view synthesis algorithm for light field images.