基于联结残差网络的单幅图像去雨

Single image deraining based on the concatenation residual network

本文通过改进的残差网络,学习有雨图像和无雨图像之间的映射关系来实现图像去雨,提出了一种基于联结残差网络的单幅图像去雨算法。首先,利用改进的残差块简化网络模块,减少网络参数,提升网络训练速度;其次,设计的联结结构不仅实现了多特征提取,有效减少了参数,而且增加了特征图的输出,有利于保留更多的图像细节信息;最后,利用单尺度卷积实现图像细节重建,提高去雨图像的视觉效果。实验结果表明本文算法在合成雨天图像数据集上,其结构相似度和峰值信噪比的平均值分别高于0.95和33 dB,而真实雨天图像数据集的盲图像质量评价值较低。本文算法不仅能有效去除图像中的雨,雨纹残留较少,而且能更多地保留图像的纹理和边缘细节,视觉效果清晰自然。

In this paper,the mapping relation between the rainy image and the clear image is learned through the improved residual network to realize the image rain removal,and a single image deraining algorithm based on the concatenation residual network is proposed.Firstly,the improved residual block is used to simplify the network module,reduce the network parameters and improve the network training speed.Secondly,the designed concatenation structure not only realizes multi-feature extraction,effectively reduces parameters,but also increases the output of feature map,which is beneficial to retain more details of the image.Finally,the single-scale convolution is used to reconstruct the image details,and further improve the visual effect of the de-rained image.Experimental results indicate that the mean values of structural similarity and peak signal-to-noise ratio on the synthetic rainy image sets are both higher than 0.95 and 33 dB,and the blind image quality evaluation values of real rainy image sets are relatively low.The method can not only remove the rain in the image effectively,the rain streak residue is less,but also keep more texture and edge details of the image,and the visual effect is clear and natural.