Combined with Pyramid Split Attention and Multi-Scale Feature Learning Network for Single Image Deraining

Abstract

The difficulty of single image deraining using deep learning method lies in how to extract feature information of rain streaks. To effectively extract discriminant information, we propose a network for single image deraining, termed dense network of combing pyramid split attention with multi-Scale feature learning (DPMNet). We utilize improved DenseNet based on pyramid split attention (PSA) module as the overall framework of DPMNet, where each dense layer is composed of a multi -scale feature learning (MSFL) module. The PSA module aims to learn important features using the attention mechanism. The MSFL module aims to learn the features of different scales, where re-parameterization VGG (Rep VGG) block is used for multi-scale feature extraction and mobile inverted bottleneck (MB) convolution is used for multi-scale feature fusion. By introducing pyramid split attention and re-parameterization VGG skillfully, the proposed method can obtain distinguishing rain features efficiently. Extensive experiments on benchmark datasets as well as specific real datasets demonstrate that the proposed method outperforms recent state-of-the-art methods.