Scholarlink

<p>Rainy weather degrades the performance of autonomous driving and video surveillance systems, emphasising the need for effective image de-raining techniques. Existing, Convolutional Neural Networks (CNNs) based methods are limited by their restricted receptive fields, while Vision Transformers (ViT) based approaches, although capable of capturing global dependencies, suffer from high computational complexity. To address these limitations, this paper proposes a Hybrid Swin Transformer Network for Single Image Deraining, termed SwinMFR-Net. The proposed model integrates hierarchical Swin Transformer blocks with a Multi-Level Feature Pyramid Fusion (MLFPF) module, combining a Feature Pyramid Network(FPN) and Multi-Scale Feature Fusion(MSFF) to effectively aggregate contextual information across scales. Additionally, Convolutional Block Attention Module (CBAM) are incorporated at multiple stages to enhance spatial and channel-wise feature representation, enabling precise rain streak removal and detail preservation. Extensive experiments conducted on multiple benchmark dataset, including Rain100H, Rain100L, Rain1400, Rain1200, Rain800 and real-world dataset, demonstrate that the proposed method consistently outperforms state-of-the-art approaches. In particular, SwinMFR-Net achieves up to 39.82dB PSNR and 0.989 SSIM on Rain100L, with significantly lower LPIPS values, while maintaining a lightweight design of approximately 12 million parameters. These results confirm the effectiveness of the proposed framework in producing high-quality, visually consistent derained images under diverse and challenging conditions.</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

Hybrid swin transformer framework for accurate and efficient single image deraining

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Rajeswari Malothu,
  • Prakash Kodali

摘要

Rainy weather degrades the performance of autonomous driving and video surveillance systems, emphasising the need for effective image de-raining techniques. Existing, Convolutional Neural Networks (CNNs) based methods are limited by their restricted receptive fields, while Vision Transformers (ViT) based approaches, although capable of capturing global dependencies, suffer from high computational complexity. To address these limitations, this paper proposes a Hybrid Swin Transformer Network for Single Image Deraining, termed SwinMFR-Net. The proposed model integrates hierarchical Swin Transformer blocks with a Multi-Level Feature Pyramid Fusion (MLFPF) module, combining a Feature Pyramid Network(FPN) and Multi-Scale Feature Fusion(MSFF) to effectively aggregate contextual information across scales. Additionally, Convolutional Block Attention Module (CBAM) are incorporated at multiple stages to enhance spatial and channel-wise feature representation, enabling precise rain streak removal and detail preservation. Extensive experiments conducted on multiple benchmark dataset, including Rain100H, Rain100L, Rain1400, Rain1200, Rain800 and real-world dataset, demonstrate that the proposed method consistently outperforms state-of-the-art approaches. In particular, SwinMFR-Net achieves up to 39.82dB PSNR and 0.989 SSIM on Rain100L, with significantly lower LPIPS values, while maintaining a lightweight design of approximately 12 million parameters. These results confirm the effectiveness of the proposed framework in producing high-quality, visually consistent derained images under diverse and challenging conditions.