Scholarlink

<p>To address the challenges of traditional peanut seed sorting—low manual efficiency, high mechanical damage rates, and insufficient identification of complex defects—an intelligent sorting device based on the RK3566 processor is designed and implemented. The system integrates a spiral vibrating feeder, industrial-camera imaging, lightweight neural-network inference, and pneumatic actuators, forming a closed-loop automated process of seed supply–imaging–recognition–decision–actuation. A lightweight convolutional neural network (LAN-SS) with channel-wise attention (SE) is developed on MobileNetV2, reducing the parameter count and model size to 0.89&#xa0;M and 3.52&#xa0;MB, respectively (39.9% and 40.3% of the original), while maintaining high accuracy. On a dataset of 3,948 manually labeled peanut seed images, the proposed LAN-SS model achieves a validation accuracy of 0.99, demonstrating that the lightweight architecture does not compromise recognition performance. Prototype test results show that the system can sort approximately 200 seeds per minute, with an average recognition accuracy of 0.942, and an average total processing time of 0.332&#xa0;s per seed (0.256&#xa0;s for image preprocessing, 0.059&#xa0;s for model inference, and 0.017&#xa0;s for other control operations), corresponding to a stable throughput of about 3 seeds per second. The system-level comprehensive evaluation metric, Correctly Handle the Success Rate of Seeds (CHSRS), reaches 0.989, confirming real-time performance and reliability for engineering deployment. This study demonstrates a novel, integrated hardware-software solution that effectively bridges the gap between high-accuracy deep learning models and their practical deployment in resource-constrained agricultural settings.</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

Intelligent sorting of peanut seeds: embedding lightweight deep learning into low-cost RK3566

  • 期刊论文

  • 发表日期：

    2026年1月30日

查看全文
  • Zhaolei Yang,
  • Xiang Li,
  • Zewen An,
  • Peng Xu,
  • Junhui Liu,
  • Nianrui Liu,
  • Yuliang Yun

摘要

To address the challenges of traditional peanut seed sorting—low manual efficiency, high mechanical damage rates, and insufficient identification of complex defects—an intelligent sorting device based on the RK3566 processor is designed and implemented. The system integrates a spiral vibrating feeder, industrial-camera imaging, lightweight neural-network inference, and pneumatic actuators, forming a closed-loop automated process of seed supply–imaging–recognition–decision–actuation. A lightweight convolutional neural network (LAN-SS) with channel-wise attention (SE) is developed on MobileNetV2, reducing the parameter count and model size to 0.89 M and 3.52 MB, respectively (39.9% and 40.3% of the original), while maintaining high accuracy. On a dataset of 3,948 manually labeled peanut seed images, the proposed LAN-SS model achieves a validation accuracy of 0.99, demonstrating that the lightweight architecture does not compromise recognition performance. Prototype test results show that the system can sort approximately 200 seeds per minute, with an average recognition accuracy of 0.942, and an average total processing time of 0.332 s per seed (0.256 s for image preprocessing, 0.059 s for model inference, and 0.017 s for other control operations), corresponding to a stable throughput of about 3 seeds per second. The system-level comprehensive evaluation metric, Correctly Handle the Success Rate of Seeds (CHSRS), reaches 0.989, confirming real-time performance and reliability for engineering deployment. This study demonstrates a novel, integrated hardware-software solution that effectively bridges the gap between high-accuracy deep learning models and their practical deployment in resource-constrained agricultural settings.