Scholarlink

<p>Deep learning has made significant strides in object detection. However, in fruit recognition for smart agriculture, existing methods often prioritise accuracy at the cost of increased computational complexity, larger parameter sizes, and slower inference. Balancing precision and efficiency remains a key challenge. To address this, we propose CHCG-YOLO, a lightweight fruit recognition model. First, CHGNetV2(Coordinate Attention HGNetV2) is introduced into the backbone, integrating HGStem and HGBlock, with a CA(Coordinate Attention) mechanism embedded in HGBlock to enhance target feature representation and reduce computational cost. Second, the CCFC(C3k2 ConvFormer ConvolutionalGLU) module is incorporated into the Neck, combining ConvFormerCGLU and lightweight C3k2 structures to boost multi-scale feature extraction, suppress background interference, and improve recognition accuracy. Third, a lightweight DEGC(Detect Group Conv) detection head is designed based on a shared parameter optimisation strategy, improving computational efficiency and simplifying the model structure. Using YOLO11n as the baseline, experiments demonstrate that CHCG-YOLO achieves a mAP50(mean Average Precision at IoU threshold 0.5) of 96.6%, while reducing parameter count and FLOPs(Floating Point Operations per Second) by 30.7% and 30.1% respectively. The model size is compressed to 4.0&#xa0;MB, and inference speed increases by 5.2%. These results demonstrate that CHCG-YOLO balances high accuracy with lightweight design. Its low computational cost and compact size make it feasible for deployment on resource-limited devices, while its improved precision and speed highlight its potential for large-scale smart agriculture.</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

CHCG-YOLO: a lightweight model for fruit recognition

  • 期刊论文

  • 发表日期：

    2025年11月22日

查看全文
  • Xinjun An,
  • Liyuan Sun,
  • Youjun Zhao,
  • Hongjuan Wang,
  • Minghui Zhang,
  • Kunxi Li,
  • Renjie Xie

摘要

Deep learning has made significant strides in object detection. However, in fruit recognition for smart agriculture, existing methods often prioritise accuracy at the cost of increased computational complexity, larger parameter sizes, and slower inference. Balancing precision and efficiency remains a key challenge. To address this, we propose CHCG-YOLO, a lightweight fruit recognition model. First, CHGNetV2(Coordinate Attention HGNetV2) is introduced into the backbone, integrating HGStem and HGBlock, with a CA(Coordinate Attention) mechanism embedded in HGBlock to enhance target feature representation and reduce computational cost. Second, the CCFC(C3k2 ConvFormer ConvolutionalGLU) module is incorporated into the Neck, combining ConvFormerCGLU and lightweight C3k2 structures to boost multi-scale feature extraction, suppress background interference, and improve recognition accuracy. Third, a lightweight DEGC(Detect Group Conv) detection head is designed based on a shared parameter optimisation strategy, improving computational efficiency and simplifying the model structure. Using YOLO11n as the baseline, experiments demonstrate that CHCG-YOLO achieves a mAP50(mean Average Precision at IoU threshold 0.5) of 96.6%, while reducing parameter count and FLOPs(Floating Point Operations per Second) by 30.7% and 30.1% respectively. The model size is compressed to 4.0 MB, and inference speed increases by 5.2%. These results demonstrate that CHCG-YOLO balances high accuracy with lightweight design. Its low computational cost and compact size make it feasible for deployment on resource-limited devices, while its improved precision and speed highlight its potential for large-scale smart agriculture.