多分辨率融合密集网络的图像去雨方法

Multi-resolution Fusion Dense Network for Image Rain Removal

针对现有的去雨方法无法彻底去除雨纹并且去雨后图像存在细节丢失问题,提出一种多分辨率融合密集网络的图像去雨方法。网络主体由多个多分辨并行融合模块构成,始终保持空间精确的高分辨率并从低分辨率中接收大量的上下文信息。使用一种基于选择性卷积核机制SKNet的多尺度特征融合模块,通过非线性的方法有效聚合来自不同分辨率流的特征。在不同的分辨率流中使用一种改进的残差模块,采用相邻层次的多种尺度的卷积来获取丰富的雨纹信息。模块间使用密集连接,加强不同模块之间的特征传播。实验表明,所提方法在合成及真实雨像数据集上的评价指标与其他去雨方法相比有所提高,去除雨纹的同时能够保留更多的细节信息。

The existing rain removal methods cannot completely remove rain streak,and the image details are lost after rain removal.Therefore,an image rain removal method based on multi-resolution fusion dense network is proposed.The main body of the network is composed of multiple multi-resolution parallel fusion modules,which always keeps the spatial accurate high-resolution representation and receives a lot of context information from the low resolution.A multi-scale feature fusion module based on the selective convolution kernel mechanism SKNet is used to effectively aggregate features from streams with different resolutions by nonlinear method.An improved residual module is used in different resolution streams,and the convolution of multiple scales of adjacent layers is used to obtain rich rain ripple information.Dense connections are used between modules to enhance feature propagation between different modules.Experiments show that the evaluation index of the proposed method on synthetic and real rain image datasets is improved compared with other rain removal methods,and more detailed information can be retained while removing rain patterns.