Scholarlink

<p>Issues such as high-frequency noise interference and loss of detailed target features due to insufficient illumination make object detection in low-light environments challenging. In particular, general-purpose detectors do not address the trade-off between noise suppression and detail recovery, resulting in reduced accuracy. To tackle these problems, this study proposes Balanced Noise Suppression and Detail Recovery YOLO (BNSDR-YOLO), designing two innovative plug-and-play modules: an Adaptive Low-pass Filter (ALF) and a Detail Enhancement Block (DEB). Specifically, the ALF proactively suppresses high-frequency noise in downsampled feature maps, while the DEB synergizes a Detail Enhancement Convolution (DEConv) and an Adaptive Fine-grained Channel Attention (FCA) mechanism to selectively recover detailed textures lost due to insufficient illumination. Experimental validations conducted on the ExDark dataset demonstrate that BNSDR-YOLO achieves a mAP@50 of 76.7% in low-light scenarios. This performance not only outperforms the baseline YOLOv11 model by 1.1% but also surpasses mainstream competing methods, such as Mamba-YOLO, by 2.8%, while maintaining a fast inference time of only 6.9 ms. The results indicate that the proposed method improves detection accuracy while meeting real-time requirements, achieving a balance between performance and efficiency.</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

Balancing noise suppression and detail recovery for efficient low-light object detection

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Zilin Liu,
  • Wenkai Tian,
  • Aitong Zhang,
  • Qingfeng Li,
  • Mengxia Hu,
  • Wenjie Qi,
  • Jiying Tuo

摘要

Issues such as high-frequency noise interference and loss of detailed target features due to insufficient illumination make object detection in low-light environments challenging. In particular, general-purpose detectors do not address the trade-off between noise suppression and detail recovery, resulting in reduced accuracy. To tackle these problems, this study proposes Balanced Noise Suppression and Detail Recovery YOLO (BNSDR-YOLO), designing two innovative plug-and-play modules: an Adaptive Low-pass Filter (ALF) and a Detail Enhancement Block (DEB). Specifically, the ALF proactively suppresses high-frequency noise in downsampled feature maps, while the DEB synergizes a Detail Enhancement Convolution (DEConv) and an Adaptive Fine-grained Channel Attention (FCA) mechanism to selectively recover detailed textures lost due to insufficient illumination. Experimental validations conducted on the ExDark dataset demonstrate that BNSDR-YOLO achieves a mAP@50 of 76.7% in low-light scenarios. This performance not only outperforms the baseline YOLOv11 model by 1.1% but also surpasses mainstream competing methods, such as Mamba-YOLO, by 2.8%, while maintaining a fast inference time of only 6.9 ms. The results indicate that the proposed method improves detection accuracy while meeting real-time requirements, achieving a balance between performance and efficiency.