Scholarlink

<p>In complex and dynamic industrial environments, conveyor belt deviation poses a serious threat to system stability and production safety. Therefore, developing a real-time and accurate belt deviation detection method is crucial to ensure the efficient and safe operation of the system. Traditional vision-based detection methods often suffer from poor robustness under strong interference conditions such as dust, fog, and uneven lighting, making them unsuitable for on-site applications. Although deep learning-based detection methods offer superior performance, challenges remain in terms of hardware deployment, inference speed, and stability under complex disturbances. To address these issues, this paper proposes an efficient instance segmentation network designed specifically for conveyor belt detection (EBSnet). Based on a channel shuffle mechanism, the network incorporates a highly efficient and lightweight backbone and neck architecture. Considering the large-scale and single-target characteristics of conveyor belts, a single-scale segmentation head is introduced to enhance target adaptability and detection efficiency. Furthermore, a structured pruning strategy is applied to compress the network, significantly reducing parameter count and computational cost while maintaining detection accuracy, thus improving its deployment performance on edge computing devices. To enhance the ability of the model to perceive belt edge contours, an edge-aware loss function is introduced, effectively improving segmentation accuracy in boundary regions. Experiments conducted on the embedded platform NVIDIA Jetson Xavier NX demonstrate that the EBSnet model requires only 4.6 GFLOPs and 1.052M parameters, achieving an AP@0.5 of 0.9913 on a custom industrial conveyor belt dataset. It also reaches inference speeds of 30 FPS and 21 FPS under single-stream and three-stream video inputs, respectively. These results validate that the proposed algorithm achieves high accuracy, low latency, and strong robustness in conveyor belt deviation monitoring under limited computational resources.</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

An efficient and lightweight network for detecting bulk material conveyor belts

  • 期刊论文

  • 发表日期：

    2025年12月18日

查看全文
  • Jie Luo,
  • Cong Mao,
  • Yongsheng Wang,
  • Zhenhuan Yao,
  • Zhuangtu Cui

摘要

In complex and dynamic industrial environments, conveyor belt deviation poses a serious threat to system stability and production safety. Therefore, developing a real-time and accurate belt deviation detection method is crucial to ensure the efficient and safe operation of the system. Traditional vision-based detection methods often suffer from poor robustness under strong interference conditions such as dust, fog, and uneven lighting, making them unsuitable for on-site applications. Although deep learning-based detection methods offer superior performance, challenges remain in terms of hardware deployment, inference speed, and stability under complex disturbances. To address these issues, this paper proposes an efficient instance segmentation network designed specifically for conveyor belt detection (EBSnet). Based on a channel shuffle mechanism, the network incorporates a highly efficient and lightweight backbone and neck architecture. Considering the large-scale and single-target characteristics of conveyor belts, a single-scale segmentation head is introduced to enhance target adaptability and detection efficiency. Furthermore, a structured pruning strategy is applied to compress the network, significantly reducing parameter count and computational cost while maintaining detection accuracy, thus improving its deployment performance on edge computing devices. To enhance the ability of the model to perceive belt edge contours, an edge-aware loss function is introduced, effectively improving segmentation accuracy in boundary regions. Experiments conducted on the embedded platform NVIDIA Jetson Xavier NX demonstrate that the EBSnet model requires only 4.6 GFLOPs and 1.052M parameters, achieving an AP@0.5 of 0.9913 on a custom industrial conveyor belt dataset. It also reaches inference speeds of 30 FPS and 21 FPS under single-stream and three-stream video inputs, respectively. These results validate that the proposed algorithm achieves high accuracy, low latency, and strong robustness in conveyor belt deviation monitoring under limited computational resources.