Scholarlink

<p>Single image deraining seeks to restore a clear scene from a single image affected by rain, a task that presents inherent challenges due to the spatially varying, anisotropic, and multiscale characteristics of rain streaks. The current convolutional neural network (CNN)-based approaches are efficient in preserving local texture but poor in long-range dependencies, whereas the Transformer-based approaches are more efficient to model global contextual reasoning at a high cost in computing time. Furthermore, most of the frameworks fail to consider explicit rain-layer models that can be of great structural help when restoring. This paper presents HybridDerainNet, which is a rain-layer guided two-stream CNN-Transformer network that consumes an estimated rain layer as auxiliary input to aid in rain-background separation. The suggested design combines fine-grained spatial preservation using residual convolutional blocks with a small self-attention block to preserve spatial fine details and to effectively combine global information in a simple U-shaped encoder-decoder structure. This design allows local–global representation learning to be balanced and it is computationally efficient. Extensive experiments on Rain1400 bench-mark showed that HybridDerainNet results in 32.78&#xa0;dB PSNR and 0.93 SSIM in only 0.75&#xa0;M parameters and 12.8G FLOPs, making a good trade-off between the quality of restoration and model complexity. These findings indicate that explicit rain-aware modeling and hybrid feature extraction is effective in the practical single image deraining.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

HybridDerainNet: A Lightweight Dual-Stream CNN-Transformer Network for Rain Layer-Aware Single Image Deraining

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • Sondip Poul Singh,
  • Nusrat Sharmin

摘要

Single image deraining seeks to restore a clear scene from a single image affected by rain, a task that presents inherent challenges due to the spatially varying, anisotropic, and multiscale characteristics of rain streaks. The current convolutional neural network (CNN)-based approaches are efficient in preserving local texture but poor in long-range dependencies, whereas the Transformer-based approaches are more efficient to model global contextual reasoning at a high cost in computing time. Furthermore, most of the frameworks fail to consider explicit rain-layer models that can be of great structural help when restoring. This paper presents HybridDerainNet, which is a rain-layer guided two-stream CNN-Transformer network that consumes an estimated rain layer as auxiliary input to aid in rain-background separation. The suggested design combines fine-grained spatial preservation using residual convolutional blocks with a small self-attention block to preserve spatial fine details and to effectively combine global information in a simple U-shaped encoder-decoder structure. This design allows local–global representation learning to be balanced and it is computationally efficient. Extensive experiments on Rain1400 bench-mark showed that HybridDerainNet results in 32.78 dB PSNR and 0.93 SSIM in only 0.75 M parameters and 12.8G FLOPs, making a good trade-off between the quality of restoration and model complexity. These findings indicate that explicit rain-aware modeling and hybrid feature extraction is effective in the practical single image deraining.