基于深度细节加强网络模型的去雨方法研究

Research on Rain Removal Method Based on Strengthening Network Model of Deep Details

在有雨的天气下,摄像机捕获的雨天图像通常成像模糊不清,能见度降低,在很大程度上干扰了计算机自动检测、识别还有目标跟踪的能力。例如,在交通事故中的责任划分和警察对犯罪分子的跟踪定位上,因为雨线对重要地理位置摄像头的干扰,导致摄像头没有办法提供清晰的图像作为证据,就会产生不良的影响。基于残差ResNet能通过改变映射范围加强深度学习、把输入分为高频细节层和基础层、通过输入的直接映射到输出加强图像特征等操作构建网络:通过频域变换来分离图像使得操作目标进一步稀疏化,把图像分成高频部分和低频部分,因为雨线几乎只存在于高频部分,所以在这里笔者只对高频部分做去除雨线的操作。笔者用压缩惩罚Squeeze-and-Excitation(SE)网络层替换掉批量归一化层(BN),把Squeeze-and-Excitation(SE)网络层加入残差网络中,这样一来可以使算法操作的图像目标值域缩小,稀疏性增强。为了验证本文方法的有效性和可行性,本文在数据集上做了大量的实验,实验结果证明了本文实现的模型去除雨线效果良好,解决了雨线残留明显或背景模糊化的不足,在实验运行速度上也超过了很多单幅图像雨线去除的算法。

In rainy weather, the image captured by the camera on rainy days is usually blurred and less visible, which greatly interferes with the ability of the computer to detect, identify and track objects automatically. For example, in terms of the division of responsibility in traffic accidents and the tracking and location of criminals by the police, the interference of the rain streaks to the cameras in important geographical locations leads to the failure of the cameras to provide clear images as evidence, which will have adverse effects. Based on the residual ResNet can strengthen the deep learning by changing the mapping range, the input is divided into high frequency detail layer and base layer, through a direct mapping of input to output to strengthen the image features such as network operating building: by frequency domain transformation to separate image makes the operation objectives further thinning, the image is divided into high frequency part and low frequency part, because of the rain streaks almost only exists in the high frequency part, so here we only to the operation of the high frequency part to remove the rain streaks. We use Squeeze-and-Excitation(SE) network layer replace batch normalized(BN), the Squeeze-and-Excitation(SE) network layer added to the residue in the network, in this way can make the operation of image target range narrowed, sparse sexual enhancement.