A Lightweight Dual Branch Fusion Network for Single Image Deraining

Transformers have shown promise in high-level vision tasks, but their direct application to low-level vision can lead to artifacts and high computational costs. To strike a balance between image quality and computational efficiency, this paper introduces a lightweight dual branch fusion network for single image deraining, termed LDFNet. LDFNet’s encoder simultaneously uses CNN and Transformer, adeptly grasping global dependencies and detailed spatial nuances. To effectively fuse features from different branches, we propose Dual Branch Feature Fusion (DBFF) to continuously fuse intermediate features from two independent branches to complement each other and extract rich features to guide image reconstruction. For a streamlined network, the CNN branch employs Ghost to minimize feature map redundancy and cut down convolutional calculations. Meanwhile, the Transformer branch leverages recursive procedures and shares weights among its blocks, deepening the network but using less GPU resources and model parameters. Inside the encoder-decoder network, reuse encoder features using Gated Skip Connection (GC) to prevent over-decoding. Benchmark dataset tests indicate that LDFNet delivers commendable results, trimming down network parameters by 31 \(\%\) .