DECTNet: A detail enhanced CNN-Transformer network for single-image deraining

Abstract

Recently, Convolutional Neural Networks (CNN) and Transformers have been widely adopted in image restoration tasks. While CNNs are highly effective at extracting local information, they struggle to capture global context. Conversely, Transformers excel at capturing global information but often face challenges in preserving spatial and structural details. To address these limitations and harness both global and local features for single-image deraining, we propose a novel approach called the Detail Enhanced CNN-Transformer Network (DECTNet). DECTNet integrates two key components: the Enhanced Residual Feature Distillation Block (ERFDB) and the Dual Attention Spatial Transformer Block (DASTB). In the ERFDB, we introduce a mixed attention mechanism, incorporating channel information-enhanced layers within the residual feature distillation structure. This design facilitates a more effective step-by-step extraction of detailed information, enabling the network to restore fine-grained image details progressively. Additionally, in the DASTB, we utilize spatial attention to refine features obtained from multi-head self-attention, while the feed-forward network leverages channel information to enhance detail preservation further. This complementary use of CNNs and Transformers allows DECTNet to balance global context understanding with detailed spatial restoration. Extensive experiments have demonstrated that DECTNet outperforms some state-of-the-art methods on single-image deraining tasks. Furthermore, our model achieves competitive results on three low-light datasets and a single-image desnowing dataset, highlighting its versatility and effectiveness across different image restoration challenges.