Scholarlink

<p>In low-light conditions, image details will be severely lost due to the lack of illumination and noise interference, resulting in unsatisfactory performance based on existing object detection methods. To address this issue, we propose a Two-Way Pyramid Enhancement Network (TPE-NET) and cascade it with YOLOv3 to construct an end-to-end low-light object detection framework, termed TPE-YOLO. First, TPE-NET uses Laplacian pyramid to divide low-light images into Gaussian low-frequency components and Laplacian high-frequency components of different scales. Specifically, for Gaussian low-frequency components, we design Low-Frequency Detail Enhancement Module (LDEM) to filter out image noise and enhance the semantic and texture information of the image. For Laplacian high-frequency components, High-Frequency Edge Enhancement Module (HEEM) is proposed to capture the edge and global feature information of the image. Secondly, Context Aggregation Module (CAM) is designed to mine and aggregate multi-scale contextual semantics through parallel dilated convolutions to further enhance the feature representation of high- and low-frequency information components. In addition, Color Extraction Module (CEM) and color consistency loss are designed to supplement color information for restored images and reduce image color distortion, thereby making object features more discriminative. We evaluate our proposed method on the low-light object detection benchmark dataset ExDark. Experimental results show that TPE-YOLO performs well under various low-light conditions, achieving 78.4% mAP, which is better than the existing low-light object detection methods. Code and weights are available at <a href="https://github.com/zzxf123/TPE-YOLO">https://github.com/zzxf123/TPE-YOLO</a>.</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

TPE-YOLO: improved low-light object detection using a two-way pyramid enhancement network

  • 期刊论文

  • 发表日期：

    2025年6月24日

查看全文
  • Xiaofei Zhang,
  • Xiaoguang Di,
  • Runwen Zhu

摘要

In low-light conditions, image details will be severely lost due to the lack of illumination and noise interference, resulting in unsatisfactory performance based on existing object detection methods. To address this issue, we propose a Two-Way Pyramid Enhancement Network (TPE-NET) and cascade it with YOLOv3 to construct an end-to-end low-light object detection framework, termed TPE-YOLO. First, TPE-NET uses Laplacian pyramid to divide low-light images into Gaussian low-frequency components and Laplacian high-frequency components of different scales. Specifically, for Gaussian low-frequency components, we design Low-Frequency Detail Enhancement Module (LDEM) to filter out image noise and enhance the semantic and texture information of the image. For Laplacian high-frequency components, High-Frequency Edge Enhancement Module (HEEM) is proposed to capture the edge and global feature information of the image. Secondly, Context Aggregation Module (CAM) is designed to mine and aggregate multi-scale contextual semantics through parallel dilated convolutions to further enhance the feature representation of high- and low-frequency information components. In addition, Color Extraction Module (CEM) and color consistency loss are designed to supplement color information for restored images and reduce image color distortion, thereby making object features more discriminative. We evaluate our proposed method on the low-light object detection benchmark dataset ExDark. Experimental results show that TPE-YOLO performs well under various low-light conditions, achieving 78.4% mAP, which is better than the existing low-light object detection methods. Code and weights are available at https://github.com/zzxf123/TPE-YOLO.