基于深度学习的视频图像再压缩方法  

Video and image recompression based on deep learning

作　　者：郝小龙 冯敏 樊强 彭启伟 韩斌 HAO Xiao-long;FENG Min;FAN Qiang;PENG Qi-wei;HAN Bin(Nanjing NARI Information Communication Technology Co. Ltd.,Nanjing 210003,China)

机构地区：[1]南京南瑞信息通信科技有限公司,江苏南京210003

出　　处：《电子设计工程》2019年第11期189-193,共5页Electronic Design Engineering

摘　　要：传统的视频图像压缩方法主要是针对视觉数据本身进行压缩,对帧间数据的压缩比例不高。本文提出一种基于深度学习技术的针对视频片段数据的压缩方法,在传统压缩方法的基础上对构成视频片段的视频帧序列进行进一步的高比例压缩。首先基于时空卷积神经网络对视频帧序列进行分类,将视频片段分为剧变、渐变和普通三个类别,再对普通类别的视频片段进行处理以提取关键帧并用关键帧表达该视频片段的信息。在重构质量相同的情况下,本文提出的方法可将H.264/AVC和HEVC编码标准的平均码率分别降低57.62%和8.37%。The traditional video and image coding algorithms mainly focus on compressing the original frames in the video or image data,whose compression rate is not significant with respect to the interframe redundancy.In this paper,we propose a deep learning based method to compress different video clips according to their specific characteristics,which could achieve a high compression rate with negligible distortion.Firstly,we classify the video clips into three categories(i.e.,drastic change;gradual change;normal change)via the spatio-temporal convolution neural network.Then,for the clips with normal change,we extract their key frames and compress them with existing standard codec.Experimental results show that the proposed method saves the bitrate by 57.62%and 8.37%with respect to the H.264/AVC and HEVC.

关 键 词：视频图像压缩 深度学习 关键帧提取 帧间冗余 

分 类 号：TN919.8[电子电信—通信与信息系统]