GPU CUDA accelerated video inpainting using synthetic algorithms

Video inpainting is essential in many multimedia applications such as film, media restoration, professional post-production and editing for personal use. Most algorithms are unfriendly to parallelization and are biased towards linear sequence processing. This paper presents a universal and automatic method for video inpainting. A heuristic method is used to avoid searching the whole patch in the case of small-scale occlusion and it also reduces the blurriness caused by limited area search in the case of large-scale occlusion. Eliminating "incorrect" patches in the initialization step of patch search, log 1 + 3 rounds propagation and optimized random search are varied to improve the performance of patch search and reconstruction. For small-scale region inpainting, some additional novel factors are added, such as the position of the occlusion, the Isophote directions of field points and the ratio of effective information. This makes the Gaussian kernel weight coefficients more accurate and more representative. For industrial use, the entire process is carried out completely on GPUs. Experiments were conducted on a series of test videos as well as samples used by other competing algorithms. The proposed method demonstrates the effectiveness of inpainting through the qualitative and quantitative results provided in this paper.