Multi-Scale Feature Fusion and Attention-Enhanced Image Deraining Network for Computer Vision
Year: 2026
Type: conference-paper
Abstract: Rain streak removal is a crucial pre-processing task in outdoor computer vision systems such as autonomous driving, surveillance, and remote sensing. Deraining approaches that are currently available do not generally capture the multi-scale structure of rain without losing fine details from the image. To address the diversity of rain streaks at different spatial scales, we present a new deep-learning-based deraining network called the Multi-Scale Feature Fusion and Attention-Enhanced Deraining Network (MSFAED-Net) in this paper. Our architecture features an encoder-decoder backbone with a special Multi-Scale Feature Fusion Module (MSFF), which aggregates feature maps learned from three spatial resolutions. Then we introduce two special attention mechanisms: Channel Attention (CA) using Squeeze and Excitation blocks, and Spatial Attention (SA) using Convolutional Block Attention Module (CBAM) to adaptively focus on informative regions. Results on benchmark datasets (Rain100H, Rain100L, and Rain1400) show state-of-the-art performance with PSNR of 40.25 dB and SSIM of 0.986 for Rain100H, which is substantially superior to all previous methods. The complementarity of each of the proposed components is confirmed by ablation studies.
Authors: Xinyi Zhu, Jun Wu, Han Wang
Language: English
Topic: Image Enhancement Techniques
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7203634735