Video Enhancement And Its GPU Implementation

With the development of high-definition digital television technology,consumer's expectations for visual perception have stimulated the demand for high-quality video resolution.Resolution is an important indicator for video content evaluation.Super-resolution make it possible to show a clearer visual effects on HD or even UHD display devices.This article focuses on high-quality video de-interlacing and super-resolution.In order to improve the quality of video de-interlacing,this paper presents a motioncompensated adaptive de-interlacing algorithm.A new texture-based intra-interpolation model is proposed to improve the accuracy of direction estimation and interpolation robustness.At the same time,the motion estimation and compensation of interlaced video is analyzed in detail.And a mechanism of motion compensation,a validity test of motion vector and a motion adaptive integration is provided.Experiments show that the proposed technique can produce more convincing de-interlaced videos with less artifacts on both subjective and objective evaluation in comparison with other methods,including the leading multimedia tool ffmpeg.Based on the A+ super-resolution algorithm,this paper gives the optimization to improve its execution speed and super-resolution quality.From the aspect of execution efficiency,a multilevel dictionary training model is proposed.It greatly reduces the execution time of training and upscaling,and makes it possible to augment training samples and dictionary scale effectively.To improve the visual quality,an A+ based super-resolution method with build-in enhancement is proposed.In this framework,the training images are pre-enhanced by an artifacts-free textureaware sharpener,and a function mapping from low-resolution patch features to its enhanced version is learned,so that the enhanced high-resolution image can be reconstructed directly during the process of upscaling without sharpening artifacts.Aiming at the demand of fast 2K to 4K video super-resolution,this paper also presents a highly efficient GPU acceleration scheme based on the super-resolution method with build-in enhancement.On one hand,the efficiency of GPU memory access is greatly improved,where both the reading redundancy and writing collision are avoided.On the other,locally shared data are preloaded to the low latency on-chip memory,and the large volume data are coalesce accessed,which effectively decreases the high latency of accessing external memory.The experiments demonstrate that the proposed 2K to 4K up-converting achieves 130ms/frame on Nvidia GTX 980 TI with enhanced visual quality.