Asymmetric Network based Dilated Convolution Transformer for Single Image Deraining
Xianhao Wu, Wanshun Si, Jindi Wu, Yingchao Song · 2023
The recent dramatic progress of transformer-based methods over convolutional neural networks (CNNs) for single image deraining is attributed to the transformer's powerful ability to model non-local messages. Indeed, rich local-global information characterization is equally important to better satisfy the derivation requirements. In this paper, we present an efficient image deraining method that integrates CNN models into the Transformer backbone to accelerate network convergence, termed Asymmetric network based Dilated convolution Transformer (ADT), which leverages Transformer's ability to learn non-local features and seamlessly integrates local detail efficiency and global structural representations. Our framework is an asymmetric architecture because the encoder focuses on shallow features contains rain shadow features, the decoder focuses on deep features, and embedding Dilconv FeedForward Networks (DFNs) within its encoder, and Dilconv self attention (DSAs) and DFNs within the decoder preserves higher similarity, resulting in high-quality image deraining. Extensive evaluation results show that our model performs superiorly and significantly improves the quality of image deraining.