Scholarlink

<p>Automobile fuel cap detection is one of the important technologies for the operation of unmanned gas stations. In order to be applicable to unstructured scenarios with high real-time requirements and easy to deploy, this paper proposes a new approach called SL-YOLO (StarNet-Lightweight-yolo) based on the basic model of YOLOv8n. First, in order to ensure lower energy consumption and latency, StarNet, which is capable of mapping inputs to a high-dimensional nonlinear feature space, is chosen in this paper to replace the backbone network of YOLOv8n. Second, in order to maximize the feature extraction efficiency of the model, the C2f-Faster module is added to the YOLOv8n neck network to replace the C2f structure. At the same time, a lightweight ESCD (Efficient Shared Convolutional Detection) head containing shared convolution is employed to achieve accurate recognition of a single small target. Finally, a LAMP-based pruning method is used to analyze and cut out the parts with lower weights. The experimental results show that the method reduces the number of parameters by 80.6% and improves the mAP50 by 0.6% compared to the original YOLOv8n model, and the final model achieves a mAP0.5 accuracy of &#xa0; 98.7%. The size of the model is only 1.28 MB, which shows good performance and can be well adapted to automotive fuel cap inspection programs.</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

Lightweight SL-YOLO algorithm for automotive fuel tank cover detection

  • 期刊论文

  • 发表日期：

    2025年4月2日

查看全文
  • Jun Dai,
  • Linghan Fu,
  • Yanqin Li,
  • Junwei Zhao,
  • Naohiko Hanajima

摘要

Automobile fuel cap detection is one of the important technologies for the operation of unmanned gas stations. In order to be applicable to unstructured scenarios with high real-time requirements and easy to deploy, this paper proposes a new approach called SL-YOLO (StarNet-Lightweight-yolo) based on the basic model of YOLOv8n. First, in order to ensure lower energy consumption and latency, StarNet, which is capable of mapping inputs to a high-dimensional nonlinear feature space, is chosen in this paper to replace the backbone network of YOLOv8n. Second, in order to maximize the feature extraction efficiency of the model, the C2f-Faster module is added to the YOLOv8n neck network to replace the C2f structure. At the same time, a lightweight ESCD (Efficient Shared Convolutional Detection) head containing shared convolution is employed to achieve accurate recognition of a single small target. Finally, a LAMP-based pruning method is used to analyze and cut out the parts with lower weights. The experimental results show that the method reduces the number of parameters by 80.6% and improves the mAP50 by 0.6% compared to the original YOLOv8n model, and the final model achieves a mAP0.5 accuracy of   98.7%. The size of the model is only 1.28 MB, which shows good performance and can be well adapted to automotive fuel cap inspection programs.