Single-image deraining is challenging due to the lack of temporal information. Current methods based on deep neural networks have achieved good performance in this task. However, these methods are still less effective in handling complex rainy situations. In this paper, we propose Channel-attention-based Multi-scale Recurrent Residual Network (CMRRNET), which tries to fully exploit contextual information of rainy images from multiple aspects. First, we construct a hybrid feature extraction module, which consists of the dilated convolution block and the multi-scale convolution block, to fully obtain image feature information. Second, we adopt the residual channel attention mechanism which makes the network aware of the importance of different channels. Third, we introduce long short-term memory to extract the correlation information of the features between different stages. We conduct extensive experiments on both synthetic and real rainy images. Ablation studies and extensive comparisons with state-of-the-art methods demonstrate the effectiveness of our CMRRNET.

由于缺乏时间信息，单图像排空是一个挑战。当前基于深度神经网络的方法在此任务中取得了良好的性能。但是，这些方法在处理复杂的下雨情况时仍然不太有效。在本文中，我们提出了基于信道注意的多尺度递归残差网络（CMRRNET），该网络试图从多个方面充分利用雨天图像的上下文信息。首先，我们构造了一个混合特征提取模块，该模块由膨胀卷积块和多尺度卷积块组成，以充分获得图像特征信息。其次，我们采用残余信道关注机制，该机制使网络意识到不同信道的重要性。第三，我们引入长短期记忆以提取不同阶段之间特征的相关信息。我们对合成和真实的雨天图像都进行了广泛的实验。消融研究以及与最新技术的广泛比较证明了我们CMRRNET的有效性。