Scholarlink

<p>To address the limitations of existing citrus grading systems, which are often bulky, expensive, and prone to causing surface damage during the grading process, we propose an automated citrus grading system based on 3D vision and deep learning. First, an automatic sorting method for scattered and stacked citrus fruits is introduced. This method uses a depth camera to guide the robot in automating the sorting of citrus placed in bins after harvesting. Additionally, a lightweight citrus surface defect detection model, CDF-DETR, is proposed. A CSCG module is designed to improve the backbone network, significantly reducing the number of parameters while enhancing detection performance. The dynamic range histogram self-attention mechanism (DHSA) was introduced to improve the AIFI module, and it accurately captures the key features by dynamically adjusting the attention points, which significantly improves the accuracy of citrus defect detection. The feature selection focusing and diffusion pyramid network (FSFD-PN) is designed to improve the feature fusion module, which effectively improves the recognition ability of the model to various damage modes. The experimental results indicate that the improved model achieves increases of 4.2% in P, 2.2% in R, and 2.8% in mAP, while reducing the number of parameters by 4.62&#xa0;M. The average grading accuracy for citrus reaches 97%, demonstrating its potential to meet practical application requirements.</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

Research on automatic citrus grading system based on 3D vision and deep learning

  • 期刊论文

  • 发表日期：

    2025年6月24日

查看全文
  • Yuewei Zhang,
  • Honglei Wei,
  • Xiuyuan Tang,
  • Yuan Gao,
  • Hao Tong

摘要

To address the limitations of existing citrus grading systems, which are often bulky, expensive, and prone to causing surface damage during the grading process, we propose an automated citrus grading system based on 3D vision and deep learning. First, an automatic sorting method for scattered and stacked citrus fruits is introduced. This method uses a depth camera to guide the robot in automating the sorting of citrus placed in bins after harvesting. Additionally, a lightweight citrus surface defect detection model, CDF-DETR, is proposed. A CSCG module is designed to improve the backbone network, significantly reducing the number of parameters while enhancing detection performance. The dynamic range histogram self-attention mechanism (DHSA) was introduced to improve the AIFI module, and it accurately captures the key features by dynamically adjusting the attention points, which significantly improves the accuracy of citrus defect detection. The feature selection focusing and diffusion pyramid network (FSFD-PN) is designed to improve the feature fusion module, which effectively improves the recognition ability of the model to various damage modes. The experimental results indicate that the improved model achieves increases of 4.2% in P, 2.2% in R, and 2.8% in mAP, while reducing the number of parameters by 4.62 M. The average grading accuracy for citrus reaches 97%, demonstrating its potential to meet practical application requirements.