基于Real ESRGAN的视频修复系统研究  

Research on Video Repair System Based on Real ESRGAN

作　　者：黄杰 夏远洋 杨晓杰 王思洁 田佩 刘涛 HUANG Jie;XIA Yuanyang;YANG Xiaojie

机构地区：[1]重庆对外经贸学院大数据与智能工程学院,重庆401520

出　　处：《科技创新与应用》2024年第15期46-49,54,共5页Technology Innovation and Application

基　　金：2022年度重庆对外经贸学院科研项目(KYKJ202201)。

摘　　要：图像修复和视频修复是计算机视觉的一项重要任务,其中图像修复又是视频修复的基础。为此,如何有效提升图像质量是实现视频质量提升的关键。传统的图像修复算法主要以样本信息为基础,通过对样本内容的扩撒来实现对破损区域的修复;由于这种方式对于图像样本有一定要求,从而制约传统图像修复技术的发展。为此,以生成新图像内容为基础的神经网络如GAN的出现,为图像修复技术转向深度学习提供方向。该课题主要以Real ESRGAN网络的图像修复技术为基础,通过对音频视频数据的隔离处理以及相同帧数据的优化和标记,构建视频修复处理流程。通过对随机视频样本的测试,并通过对单帧图片质量和视频数据流畅性与协调性的评估,该视频处理方法表现出较好的系统性能。Image repair and video repair is an important task of computer vision,in which image repair is the basis of video repair.Therefore,how to effectively improve the image quality is the key to improve the video quality.The traditional image restoration algorithm is mainly based on the sample information,through the expansion of the sample content to achieve the repair of the damaged area;because this method has certain requirements for image samples,which restricts the development of traditional image restoration technology.For this reason,the emergence of neural networks based on generating new image content,such as GAN,provides a direction for image restoration technology to turn to deep learning.This topic is mainly based on the image restoration technology of Real ESRGAN network.Through the isolation processing of audio and video data and the optimization and marking of the same frame data,the video restoration process is constructed.Through the test of random video samples and the evaluation of single-frame picture quality and video data fluency and coordination,the video processing method shows good system performance.

关 键 词：图像分割 神经网络 视频修复 Real ESRGAN 图像修复 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]