基于多尺度深层网络的监控图像去雨研究

Research on Rain Removal Method of Transmission and Distribution Equipment Image Based on Multi-scale Deep Network

雨天条件下视频监测图像中含有雨滴,使得输变配电设备监测目标的细节信息模糊。针对该问题,提出一种基于深层神经网络的多尺度模型架构,用于去除图像中的雨纹。利用先验信息,采用导向滤波提取表征图像高频分量的雨滴模糊特征图将模型聚焦于雨滴信息。借鉴Inception网络多分支提取多阶特征结构,构建多尺度深层神经网络融合底层和高层特征。在合成的输变配电设备以及真实世界的雨天图像集上,对本文方法进行实验验证,结果表明本文方法较其他方法具有更好的去雨效果。

Under rainy conditions,the video monitoring image contains raindrops,which blurs the detailed information of the monitoring target of the power substation equipment,and affects the performance of video monitoring.Aimed at the problem of image degradation caused by raindrops,a multi-scale model architecture is proposed based on deep neural networks to remove rain patterns in images.Using the prior information,the guided filtering is used to extract the fuzzy feature maps of raindrops that characterize the high-frequency components of the image,which renders the model focuses on the raindrop information.Learning from the multi-branch extraction structure of the multi-level feature of Inception network,a multi-scale deep neural network is built to fuse the bottom and high-level features.On the synthetic power substation equipment and real-world rainy day image sets,the method proposed has been experimentally verified.The experimental results show that the method has better rain removal effect than other methods.