Research On A Variational Method For Making Binocular Stereo Images

In recent years, the stereo display technology has made huge progress, applications such as stereo camera, 3D movie and 3D television have been gaining their ground in our daily life. Compared with the rapid development of stereo display technology, the technology of stereo content generation falls behind relatively. View synthesis is one of the key technology of stereo generation. Existing stereo generation methods based on view synthesis can’t utilize the existing views effectively, and may introduce extra noise in the image inpainting process. In addition, because traditional view synthesis methods can’t specify viewpoint for each pixel of the synthesized view, it is hard to realize per-pixel disparity control.For the problems existing in applying traditional view synthesis methods to stereo generation, a method based on variational optimization is proposed to synthesize the view by per-pixel viewpoint control. This method expresses view synthesis as a variational energy minimization problem, and defines a per-pixel viewpoint energy term. By minimizing the energy term, we make the synthesized view approach to the target view to satisfy the constraints. Meanwhile, we significantly reduce the computational complexity by an explicitly weighted average calculation. The view synthesis experiments based on light field images demonstrate the validity and effectiveness of our method. Compared with traditional methods, the PSNR values of synthesized views are generally improved by 3%.Based on the variational view synthesis method above, a variational method for making binocular stereo images/videos is further proposed. This method integrates view synthesis and image inpainting process into a variational energy optimization process which can reduce the artifacts caused by the image inpainting and utilize more than two input views effectively. This method changes the depth impression of the scene locally by per-pixel disparity mapping. Firstly we compute the per-pixel viewpoint positions of the target view according to the target disparity. Then we use the synthesized novel view to generate stereo images/videos which satisfy the disparity constraints. Stereo generation experiments based on light filed images and Kinect videos demonstrate the effectiveness and usability of our method. Compared with traditional methods, our method achieves a nearly 50% improvements both in time and memory consumption.