2018 IEEE 25th Symposium on Computer Arithmetic (ARITH)

1 / 2017-10-20 17:28:15

GPU CUDA Accelerated Video Inpainting using Synthetic Algorithms

GPU, isotropic diffusion model, parallel computing, patch match, synthetic algo-rithm, video inpainting

Video inpainting is essential in many multimedia applications such as film, media restoration, professional post-production and editing by personal use. The fea-tures of the images and videos aren't fully utilized in most existing algorithms and most algorithms are unfriendly parallel for linear sequence processing. This paper presents a unified and automatic method for video inpainting. We use synthetic method to avoid searching the whole patches with small-scale region and blurri-ness caused by limited area search in large-scale occlusion. We eliminate appar-ently "incorrect" patches in initialization step of patch search. The log⁡l+3 rounds propagation, random search integrated with times and maximum search radius of current status and texture pyramid can improve the performance of patch search and reconstruction. For small-scale region inpainting, we add some additional factors, such as the position of the occlusion, the Isophote directions of field points and ratio of effective information. Those make the Gaussian kernel weight coefficients more accurate and more characteristic. For industrial use, the entire process flow is done completely on GPU. Experiments were conducted with both a series of testing videos and samples used by other algorithms. The proposed method demonstrated the effective capacity for inpainting through the qualitative and quantitative results.