Reconstruction of Multi-view Video Based on GAN

There is a huge amount of data in multi-view video which brings enormous challenges to the compression, storage, and transmission of video data. Transmitting part of the viewpoint information is a prior solution to reconstruct the original multi-viewpoint information. They are all based on pixel matching to obtain the correlation between adjacent viewpoint images. However, pixels cannot express the invariability of image features and are susceptible to noise. Therefore, in order to overcome the above problems, the VGG network is used to extract the high-dimensional features between the images, indicating the relevance of the adjacent images. The GAN is further used to more accurately generate virtual viewpoint images. We extract the lines at the same positions of the viewpoints as local areas for image merging and input the local images into the network. In the reconstruction viewpoint, we generate a local image of a dense viewpoint through the GAN network. Experiments on multiple test sequences show that the proposed method has a 0.2-0.8-dB PSNR and 0.15-0.61 MOS improvement over the traditional method.