Scholarlink

<p>Automated screw detection is a critical yet challenging task for the disassembly of electric vehicle (EV) batteries, a process hampered by visually complex industrial environments and a lack of public datasets. This study provides the first systematic benchmark comparing two leading object detection paradigms: YOLOv11l (a one-stage detector) and Faster Region-based Convolutional Neural Network (CNN, a two-stage detector). Using a combination of a proxy dataset (laptop screws) and a custom-collected EV screw dataset, we evaluate the models across three scenarios: baseline performance, robustness to controlled visual perturbations (blur, brightness, contrast), and domain-specific fine-tuning. Our 5-fold cross-validation results show that while Faster R-CNN achieves statistically superior accuracy on the fine-tuned test set, YOLOv11l demonstrates significantly greater robustness to variations in lighting and contrast. We conclude that a clear trade-off exists: Faster R-CNN is optimal for controlled, high-accuracy offline inspection, whereas YOLOv11l’s resilience makes it the more viable candidate for real-time robotic deployment in variable factory settings. This work establishes a crucial empirical foundation for selecting and deploying robust computer vision systems in the growing field of automated battery recycling.</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

Comparative analysis of YOLO and faster R-CNN models for EV battery screw detection

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Alireza Naseri,
  • Syad Khwajazada,
  • Sheng Yang

摘要

Automated screw detection is a critical yet challenging task for the disassembly of electric vehicle (EV) batteries, a process hampered by visually complex industrial environments and a lack of public datasets. This study provides the first systematic benchmark comparing two leading object detection paradigms: YOLOv11l (a one-stage detector) and Faster Region-based Convolutional Neural Network (CNN, a two-stage detector). Using a combination of a proxy dataset (laptop screws) and a custom-collected EV screw dataset, we evaluate the models across three scenarios: baseline performance, robustness to controlled visual perturbations (blur, brightness, contrast), and domain-specific fine-tuning. Our 5-fold cross-validation results show that while Faster R-CNN achieves statistically superior accuracy on the fine-tuned test set, YOLOv11l demonstrates significantly greater robustness to variations in lighting and contrast. We conclude that a clear trade-off exists: Faster R-CNN is optimal for controlled, high-accuracy offline inspection, whereas YOLOv11l’s resilience makes it the more viable candidate for real-time robotic deployment in variable factory settings. This work establishes a crucial empirical foundation for selecting and deploying robust computer vision systems in the growing field of automated battery recycling.