基于可变形卷积的侦察视频增强方法  

A Reconnaissance Video Enhancement Method Based on Deformable Convolution

作　　者：赵彦杰 崔海斌[2] 陈振[1] 谌德荣[1] 宫久路[1] ZHAO Yanjie;CUI Haibin;CHEN Zhen;CHEN Derong;GONG Jiulu(Beijing Institute of Technology,Beijing 100081,China;Unit 91515 of PLA,Sanya 572099,China)

机构地区：[1]北京理工大学,北京100081 [2]中国人民解放军91515部队,海南三亚572099

出　　处：《探测与控制学报》2022年第5期46-52,共7页Journal of Detection & Control

摘　　要：针对现有压缩视频增强方法没有充分利用相邻视频帧间的时空相关性,造成压缩视频增强效果不明显的问题,提出基于可变形卷积的侦察视频增强方法。该方法为了排除其他视频帧的干扰,准确地捕捉视频帧在不同尺度下的运动特征,采用编码器解码器结构的分组预测网络进行多次上采样和下采样;为了适应运动目标的位移和几何变形,有效地提取相邻视频帧间的时空相关性,采用可变形卷积操作;为了充分利用相邻视频帧间的时空相关性,重建出高质量的视频帧,采用密集连接的增强网络。实验结果表明,该方法的增强效果在PSNR和SSIM指标上要优于其他经典的压缩视频增强方法。Aiming at the problem that the existing compressed video enhancement methods can not make full use of the spatio-temporal correlation between adjacent video frames,which will cause the insignificant compressed video enhancement effect,a reconnaissance video enhancement method based on deformable convolution was proposed.In order to eliminate the interference of other video frames and accurately capture the motion characteristics of video frames at different scales,this method used a group prediction network with an encoder-decoder structure to perform multiple upsampling and downsampling.For the moving targets displacement and geometric deformation,the spatio-temporal correlation was extracted effectively between adjacent video frames by using deformable convolution operation.In order to make full use of the spatio-temporal correlation between adjacent video frames,high-quality video frames were reconstructed by adopting enhanced networks with dense connections.Experimental results showed that the enhancement effect was better than other classic compression video enhancement methods in terms of PSNR and SSIM.

关 键 词：视频增强 分组预测 可变形卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]