基于生成对抗网络的图像视频编码综述

Review on image and video coding via generative adversarial networks

图像视频编码是多媒体信号处理中重点研究的问题之一,旨在高效、紧凑地表达数据,同时最大程度降低编码失真,节省传输与存储成本。经典的图像视频编码技术自上世纪七十年代起形成基于块的“预测-变换-熵编码”的混合编码框架,每一步均需要人工设计算法分别进行优化,实现像素级别的保真,然而其在低码率下由于量化丢失掉大量高频信息,会产生模糊、块效应等令人无法接受的压缩失真。近年来,基于生成对抗网络的图像视频编码的研究取得了较大的进展。相比经典方法,生成对抗网络在低码率下能够较好地弥补高频纹理细节。本文系统地梳理了基于生成对抗网络的图像视频编码的技术和进展,分别从基于全神经网络的端到端编码、生成对抗网络、基于生成对抗网络的图像视频编码三个方面进行了综述介绍,同时对基于生成对抗网络的图像视频编码的未来发展趋势进行了分析与展望。

Image and video coding is a primary research field in multimedia signal processing,whose objective is to efficiently and compactly represent data while reducing coding distortion and reducing transmission and storage costs.Traditional image video coding technology has developed a block-based hybrid"prediction-transform-entropy"coding framework which optimizes each step separately to achieve pixel-level fidelity.Quantization,however,loses a significant amount of high-frequency information at low bit rates,resulting in blurring,block effects,and other unacceptable compression distortions.A significant amount of progress has been made in recent years in the study of generative adversarial networks(GANs)for video and image coding.Compared with classical methods,GANs are able to compensate for high-frequency texture details at low bit rates.In this paper,the authors review the progress made in end-to-end coding using neural networks and GANs,and the techniques and progress associated with image video coding using GANs.Future growth trends are also assessed and forecasted.