Scholarlink

<p>This paper presents a novel lightweight Enhanced Connection Shared Network (ECSNet), based on YOLOv8n and designed specifically for efficient traffic sign detection. Motivated by the demand for both speed and precision in real-time scenarios, ECSNet incorporates two novel components: the Enhanced Connection Path Aggregation Network (ECPANet) and the Shared Convolution Head (SCHead). The ECPANet module simplifies and unifies the number of channels across various resolutions and relays shallow information to deeper layers through additional connections, effectively mitigating the loss of shallow features and enhancing feature integration, thereby reducing the number of model parameters and computational complexity. The SCHead module adapts the detection heads to different resolutions by sharing convolutions and balancing their trainable parameters, thereby markedly reducing the overall model size without sacrificing detection accuracy. Together, these innovations form a lightweight, optimized framework. Experiments on benchmark datasets demonstrate that, compared with YOLOv8n, ECSNet improves mAP by 1.69%, cuts FLOPs by 1.5G, and retains only 43.56% of the original parameters. These results highlight that ECSNet has significantly reduced the number of parameters while improving detection accuracy, making it highly suitable for real-time traffic sign detection and deployment on resource-limited devices.</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

ECSNet: A Lightweight and Enhanced YOLOv8–Based Model for Traffic Sign Detection

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Yifan Ouyang,
  • Yingqian Zhang

摘要

This paper presents a novel lightweight Enhanced Connection Shared Network (ECSNet), based on YOLOv8n and designed specifically for efficient traffic sign detection. Motivated by the demand for both speed and precision in real-time scenarios, ECSNet incorporates two novel components: the Enhanced Connection Path Aggregation Network (ECPANet) and the Shared Convolution Head (SCHead). The ECPANet module simplifies and unifies the number of channels across various resolutions and relays shallow information to deeper layers through additional connections, effectively mitigating the loss of shallow features and enhancing feature integration, thereby reducing the number of model parameters and computational complexity. The SCHead module adapts the detection heads to different resolutions by sharing convolutions and balancing their trainable parameters, thereby markedly reducing the overall model size without sacrificing detection accuracy. Together, these innovations form a lightweight, optimized framework. Experiments on benchmark datasets demonstrate that, compared with YOLOv8n, ECSNet improves mAP by 1.69%, cuts FLOPs by 1.5G, and retains only 43.56% of the original parameters. These results highlight that ECSNet has significantly reduced the number of parameters while improving detection accuracy, making it highly suitable for real-time traffic sign detection and deployment on resource-limited devices.