基于双目视差机制的视频图像超分辨率重建  

Super-resolution Reconstruction of Video Image Based on Binocular Parallax Mechanism

作　　者：郝贵举 杨洪臣[1] 颜瑞彬 HAO Guiju;YANG Hongchen;YAN Ruibin(Criminal Investigation Police University of China,Shenyang 110035,China;People’s Public Security University of China,Beijing 100038,China)

机构地区：[1]中国刑事警察学院,沈阳110035 [2]中国人民公安大学,北京100038

出　　处：《刑事技术》2022年第5期458-464,共7页Forensic Science and Technology

基　　金：上海市现场物证重点实验室开放课题(2020XCWZK03)。

摘　　要：针对公安实践中成像质量较差的视频图像,基于双目视差机制进行超分辨重建。本文提出用视频中相近时段的两帧图像代替双目图像,即将摄像机位置固定,物体连续运动形成视差不同的视频图像,截取视频中同一运动物体不同时间的两帧图像作为双目图像。采用双目视差机制以及残差网络模型实现视频图像的超分辨率重建。通过多尺度特征提取、视差注意模块、超分辨率重建模块,获得图像的高频信息。结果表明,相较于传统方法,本文方法峰值信噪比提升1.27dB,结构相似性指标提升0.004,主观视觉效果上可以观察出视频图像明显清晰,减少了边缘模糊。基于双目视差机制的视频图像超分辨率重建的方法视觉效果上增加了清晰度,超分辨效果良好。Due to the influence of photograph-taking conditions including distance,competence of monitoring equipment and blurring from the moving objects,the relevant video images are frequently rendered into poor visual quality,making them difficult to either recognize or meet the investigative requirements for public security.Therefore,a super-resolution reconstruction is here proposed about video image based on binocular parallax mechanism.Particularly,with the camera fixed at a position,the tested moving object is to have its continuous motion kept into images of different parallax against which two time-different frames of the object’s picture are placed as binocular images into the adopted network model.The image is extracted of its multi-scale information with the feature extraction module outfitting improved atrous spatial pyramid pooling,followed to have its effective information obtained through introduction into the parallax attention module and successively its super-resolution right/left image(frame)reconstructed via the residual network.Finally,the deconvolution is operated to visualize the convolution process.The experimental results show that the output image,compared with the resultant from traditional video image super-resolution reconstruction,has increased its peak’s signal-to-noise ratio by 1.27 db,enhanced its structural similarity index by 0.004,with its subjective visual effect being obviously clear and its edge blur reduced,hence demonstrating that the excellence is present of the here-adopted video image super-resolution reconstruction based on binocular parallax mechanism.A new idea is thus provided for improving video image’s resolution with such a manipulation of replacing binocular images with two frames taken at similar-yet-different time points in a video and combination of the convolutional neural networks for super-resolution.The approach adopted here can also be applied to recover low-resolution video images exampled as those handled in this article about recognition of one person�

关 键 词：视频侦查 视差机制 超分辨率 双目图像 残差网络 

分 类 号：DF793.2[政治法律—诉讼法学]