基于校园监控的多帧图像超分辨率重建技术
Multi Frame Image Super-resolution Reconstruction Technology Based on Campus Monitoring
陈亮 1梁暄浩1
作者信息
- 1. 沈阳理工大学 自动化与电气工程学院,沈阳 110159
- 折叠
摘要
为解决校园监控视频质量低、细节信息少的问题,在BasicVSR网络模型的基础上提出一种增加辅助机制的多帧图像超分辨率重建技术.首先,选择双向循环机制的BasicVSR网络以保证信息传递的多样性与连贯性;其次,在图像配准模块增加辅助机制以获得关键帧特征,矫正主干网络的特征图,该辅助机制在一定程度上解决了图像长期传递的信息误差,减少了纹理造假现象;最后,使用早期图像融合的方式,将前向分支与后向分支的特征图像融合为新的特征并使用亚像素卷积进行图像重建.在自制校园监控数据集上的对比实验结果表明,改进后模型的峰值信噪比相比于原模型提升了 1.51 dB,提高了监控视频的清晰度,具有较好的应用价值.
Abstract
To solve the problem of low quality and insufficient detail information in campus sur-veillance videos,a multi frame image super-resolution reconstruction technique with added auxiliary mechanisms is proposed based on the BasicVSR network model.Firstly,the BasicVSR network with a bidirectional loop mechanism is chosen to ensure the diversity and coherence of information trans-mission.Secondly,an auxiliary mechanism is added to the image registration module to obtain key-frame features and correct the feature maps of the backbone network.This auxiliary mechanism to some extent solves the information error in long-term image transmission and reduces texture falsifi-cation.Finally,using early image fusion methods,the feature images of the forward and backward branches are fused into new features and reconstructed using sub-pixel convolution.The compara-tive experimental results of a self-made campus monitoring dataset show that the improved model has a 1.51 dB increase in peak signal-to-noise ratio compared to the original model,which im-proves the clarity of monitoring videos and has good application value.
关键词
校园监控/超分辨率重建技术/图像配准/图像融合Key words
campus monitoring/super-resolution reconstruction technology/image registration/im-age fusion引用本文复制引用
基金项目
国家重点研发计划资助项目(2017YFC0821001)
国家重点研发计划资助项目(2017YFC0821004)
出版年
2024
NETL
NSTL
万方数据