基于多路光流信息的微光视频增强算法

Low-light video enhancement algorithm based on multi-channel optical flow information

图像和视频是记录真实场景信息的重要媒介,它们包含丰富而详细的视觉内容,可以开发各种智能系统来执行各种任务。特别是对于低照度条件下的视频,提升其清晰度和细节可以更好地表现和还原真实场景。针对在夜间低照度环境条件下对周围环境感知的需求,提出一种基于多路光流信息时间一致性的微光视频增强算法。通过引入预测的光流与真实的光流信息,构建三分支孪生网络对微光视频进行增强;同时针对微光视频存在的低信噪比以及模糊化问题,设计一种基于双尺度注意力机制的微光视频去噪模块(CA-Swin模块),以提升网络的去噪性能。通过在DAVIS数据集上进行对比实验和评估,得出所提网络在增强微光视频方面更高效,鲁棒性显著;且该策略还具有通用性,可以直接扩展到大规模数据集。

Images and videos are important media for recording real scene information,which containing rich and detailed visual content,can develop various intelligent systems to perform various tasks.Especially for videos under low lighting conditions.Improving their clarity and details can better represent and restore real scenes.A low-light video enhancement algorithm based on temporal consistency of multi-channel optical flow information is proposed to meet the demand for perceiving the surrounding environment under low illumination conditions at night.The predicted optical flow and the real optical flow information are introduced to construct a three-branch twin network,so as to enhance the low-light video.In allusion to the low signal-to-noise ratio and blurring in low-light video,a low-light video denoising module(CA-Swin module)based on dual scale attention mechanism is designed to improve the denoising performance of the network.By the comparative experiment and evaluation on the DAVIS dataset,it is found that the proposed network is more efficient and has significant robustness in enhancing low-light video.This strategy also has universality and can be directly extended to large-scale datasets.