Scholarlink

<p>Lightweight convolutional neural networks have created new opportunities for object recognition, enabling high-performance algorithms to operate on resource-constrained devices while preserving strong representational and generalization capabilities. This paper introduces a lightweight YOLO-Fast object detection model. By replacing the backbone feature extraction network, redundant computations are minimized, thereby optimizing inference speed. An attention module is incorporated to enhance detection accuracy and feature fusion capability. The experimental findings based on the publicly accessible VisDrone dataset indicate that the proposed model has a 1.03% improvement in mAP over the baseline model. Through sparse training and channel pruning, the model’s inference speed is further optimized, while the number of network parameters is reduced. Moreover, knowledge distillation restores the network’s accuracy to pre-pruning levels. The model’s computational load is reduced by 35.5%, with a processing time of 12 ms per image on the embedded Atlas 200I development board. The code is available at: <a href="https://github.com/ZJ-Song-Lab/YOLO-Fast">https://github.com/ZJ-Song-Lab/YOLO-Fast</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

YOLO-Fast: a lightweight object detection model for edge devices

  • 期刊论文

  • 发表日期：

    2025年4月10日

查看全文
  • Zijing Song,
  • Xiaoyu Zhang,
  • Panlong Tan

摘要

Lightweight convolutional neural networks have created new opportunities for object recognition, enabling high-performance algorithms to operate on resource-constrained devices while preserving strong representational and generalization capabilities. This paper introduces a lightweight YOLO-Fast object detection model. By replacing the backbone feature extraction network, redundant computations are minimized, thereby optimizing inference speed. An attention module is incorporated to enhance detection accuracy and feature fusion capability. The experimental findings based on the publicly accessible VisDrone dataset indicate that the proposed model has a 1.03% improvement in mAP over the baseline model. Through sparse training and channel pruning, the model’s inference speed is further optimized, while the number of network parameters is reduced. Moreover, knowledge distillation restores the network’s accuracy to pre-pruning levels. The model’s computational load is reduced by 35.5%, with a processing time of 12 ms per image on the embedded Atlas 200I development board. The code is available at: https://github.com/ZJ-Song-Lab/YOLO-Fast.