Scholarlink

<p>To address the challenges of labor-intensive harvesting, high operational costs, and low recognition accuracy in blueberry production in Yunnan Province, this study proposes an improved YOLOv7tiny-based model (YOLOv7tiny-DGS) for precise ripeness detection. The model integrates DCNv2 into the backbone for enhanced feature extraction, incorporates GCnet to improve global context modeling, employs bilinear interpolation in the neck for refined feature fusion, and replaces CIoU with SIoU loss in the prediction head to improve bounding-box regression across scales. Trained on 6,180 blueberry images, the model was validated through ablation experiments, confirming the effectiveness of these improvements. Grad-CAM visualization further demonstrated that YOLOv7tiny-DGS exhibited stronger attention to blueberry fruit targets. Compared with eight benchmark models, YOLOv7tiny-DGS achieved gains of 2.0%–8.6% in mAP@0.5 and 2.4%–14.5% in mAP, while outperforming the baseline YOLOv7tiny by 2.5% and 2.8%, respectively. In addition, a regional ripeness detection system was developed using ONNX Runtime and a PyQt5-based GUI, demonstrating the feasibility of practical deployment. These results provide a theoretical and technical basis for optimizing automated harvest scheduling and improving operational efficiency in blueberry cultivation.</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

Blueberry regional ripeness detection based on an improved YOLOv7tiny model

  • 期刊论文

  • 发表日期：

    2025年11月9日

查看全文
  • Jianian Li,
  • Jiangfeng He,
  • Chao Wang,
  • Shuai Tan,
  • Jiazhen Zhao

摘要

To address the challenges of labor-intensive harvesting, high operational costs, and low recognition accuracy in blueberry production in Yunnan Province, this study proposes an improved YOLOv7tiny-based model (YOLOv7tiny-DGS) for precise ripeness detection. The model integrates DCNv2 into the backbone for enhanced feature extraction, incorporates GCnet to improve global context modeling, employs bilinear interpolation in the neck for refined feature fusion, and replaces CIoU with SIoU loss in the prediction head to improve bounding-box regression across scales. Trained on 6,180 blueberry images, the model was validated through ablation experiments, confirming the effectiveness of these improvements. Grad-CAM visualization further demonstrated that YOLOv7tiny-DGS exhibited stronger attention to blueberry fruit targets. Compared with eight benchmark models, YOLOv7tiny-DGS achieved gains of 2.0%–8.6% in mAP@0.5 and 2.4%–14.5% in mAP, while outperforming the baseline YOLOv7tiny by 2.5% and 2.8%, respectively. In addition, a regional ripeness detection system was developed using ONNX Runtime and a PyQt5-based GUI, demonstrating the feasibility of practical deployment. These results provide a theoretical and technical basis for optimizing automated harvest scheduling and improving operational efficiency in blueberry cultivation.