Scholarlink

<p>Image restoration plays a critical role in pattern recognition by enhancing data quality for downstream tasks such as object detection and segmentation. It aims to recover high-quality images from severely degraded inputs caused by various factors. Traditional approaches often address specific issues like low-light enhancement, dehazing, or denoising separately, which limits their practicality and generalizability in real-world applications. In this paper, we propose an efficient multitask image restoration baseline that unifies low-light enhancement, raindrop removal, and shadow removal within a single framework. Our method introduces a U-shaped neural architecture equipped with a Gated Dual-branch Attention (GatedDA) module for degradation-aware feature refinement, a Dual Attention Block (DAB) for robust feature extraction, and a cross-level feature distillation mechanism to promote knowledge transfer between network stages. This unified design enables a single model to handle multiple degradation types efficiently. Experimental results show that our model achieves a PSNR of 25.04, SSIM of 0.76, mAP@50 of 0.67 in detection, and 0.60 in segmentation while using 8 times fewer parameters than DA-CLIP, demonstrating both strong restoration quality and practical utility for pattern recognition tasks.</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

Efficient multitask model for image restoration in degraded conditions

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Duc Manh Dao,
  • Binh An Nguyen,
  • Tung Luong Nguyen,
  • Namal Rathnayake,
  • Yukinobu Hoshino,
  • Tuan Linh Dang

摘要

Image restoration plays a critical role in pattern recognition by enhancing data quality for downstream tasks such as object detection and segmentation. It aims to recover high-quality images from severely degraded inputs caused by various factors. Traditional approaches often address specific issues like low-light enhancement, dehazing, or denoising separately, which limits their practicality and generalizability in real-world applications. In this paper, we propose an efficient multitask image restoration baseline that unifies low-light enhancement, raindrop removal, and shadow removal within a single framework. Our method introduces a U-shaped neural architecture equipped with a Gated Dual-branch Attention (GatedDA) module for degradation-aware feature refinement, a Dual Attention Block (DAB) for robust feature extraction, and a cross-level feature distillation mechanism to promote knowledge transfer between network stages. This unified design enables a single model to handle multiple degradation types efficiently. Experimental results show that our model achieves a PSNR of 25.04, SSIM of 0.76, mAP@50 of 0.67 in detection, and 0.60 in segmentation while using 8 times fewer parameters than DA-CLIP, demonstrating both strong restoration quality and practical utility for pattern recognition tasks.