Scholarlink

<p>In the last few years, object detection models using deep neural network have achieved remarkable success in clear weather conditions. Nevertheless, detection performance notably declines in low-light environments, where reduced illumination severely limits feature visibility and degrades detection accuracy. To overcome this challenge, we introduce a novel Enhanced-to-Detect Network (E2D-Net), specifically designed to boost object detection performance under nighttime conditions. The architecture of E2D-Net comprises two primary subnetworks: a Low-Light Enhancement (LL) subnetwork for generating enhanced features, and an Object Detection (OD) subnetwork for classifying and localizing objects. By sharing shallow layers for multi-task learning and leveraging multiscale knowledge transmission in deeper layers, the proposed model effectively transfers enhanced features from the LL subnetwork to the OD subnetwork, thereby improving detection performance. Experimental results confirm that our E2D-Net achieves a significant improvement in classifying and localizing objects, surpassing both modern object detectors and enhancement-detection models under synthetic and real-world nighttime 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

Enhance-to-detect: a unified deep learning network for object detection in low-light environments

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Trung-Hieu Le,
  • Do-Thu-Ha Tran,
  • Quoc-Viet Hoang,
  • Ngoc-Thang Pham

摘要

In the last few years, object detection models using deep neural network have achieved remarkable success in clear weather conditions. Nevertheless, detection performance notably declines in low-light environments, where reduced illumination severely limits feature visibility and degrades detection accuracy. To overcome this challenge, we introduce a novel Enhanced-to-Detect Network (E2D-Net), specifically designed to boost object detection performance under nighttime conditions. The architecture of E2D-Net comprises two primary subnetworks: a Low-Light Enhancement (LL) subnetwork for generating enhanced features, and an Object Detection (OD) subnetwork for classifying and localizing objects. By sharing shallow layers for multi-task learning and leveraging multiscale knowledge transmission in deeper layers, the proposed model effectively transfers enhanced features from the LL subnetwork to the OD subnetwork, thereby improving detection performance. Experimental results confirm that our E2D-Net achieves a significant improvement in classifying and localizing objects, surpassing both modern object detectors and enhancement-detection models under synthetic and real-world nighttime conditions.