Scholarlink

<p>Sugarcane is the largest cash crop globally in terms of production and serves as a primary source of sugar and biofuel. The buds naturally present on sugarcane stems are used for vegetative propagation, and determining the health of the sugarcane stems is a crucial preliminary step in automated bud extraction pipelines to achieve better seed quality and higher crop yield. However, automated sugarcane stem health detection is not addressed in prior research. Additionally, agricultural automation models must be extremely lightweight and efficient for edge and embedded deployment on resource-constrained devices, where both model size and inference latency are important considerations. In this paper, we present a new lightweight object detection model, RepNCSP-ELAN-YOLOv8n (Repeated Normalised Cross Stage Partial with Efficient Layer Aggregation Network-You Only Look Once version 8 nano), for real-time sugarcane stem health classification (Fit vs. Unfit). The proposed model achieves ultra-low computational complexity of only 7.2 GFLOPs, 2.2 million parameters, and 4.6&#xa0;MB model size (in PyTorch)—significantly lower than prior YOLOv8n and all other nano-level YOLO versions—while still providing state-of-the-art performance of 90.1% mAP@50, 64.2% mAP@95, and 14.33&#xa0;ms average inference time. The model is trained on a self-curated dataset of 3839 images collected under various natural farm conditions in multiple Indian regions. We conduct extensive benchmarking with major YOLO versions, demonstrating our proposed model’s superior efficiency and competitive accuracy. Furthermore, we empirically validate the model’s real-world deployability across multiple edge AI platforms (Jetson Nano, Raspberry Pi 4B) and inference formats (PyTorch, TorchScript, ONNX, TensorRT, MNN and NCNN). The proposed model offers a real-time framework for automated sugarcane stem health detection, providing a crucial step toward intelligent, edge-enabled agricultural automation solutions for better seed quality and higher crop productivity. Source code: <a href="https://github.com/NCU-Pushpendra/project-code/blob/main/real-time-sugarcane-detection.ipynb">https://github.com/NCU-Pushpendra/project-code/blob/main/real-time-sugarcane-detection.ipynb</a>.</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

Real time sugarcane stem health detection with YOLOv8 tinier and RepNCSP ELAN

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月12日

查看全文
  • Pushpendra Kumar,
  • Shraddha Arora,
  • Shaveta Arora

摘要

Sugarcane is the largest cash crop globally in terms of production and serves as a primary source of sugar and biofuel. The buds naturally present on sugarcane stems are used for vegetative propagation, and determining the health of the sugarcane stems is a crucial preliminary step in automated bud extraction pipelines to achieve better seed quality and higher crop yield. However, automated sugarcane stem health detection is not addressed in prior research. Additionally, agricultural automation models must be extremely lightweight and efficient for edge and embedded deployment on resource-constrained devices, where both model size and inference latency are important considerations. In this paper, we present a new lightweight object detection model, RepNCSP-ELAN-YOLOv8n (Repeated Normalised Cross Stage Partial with Efficient Layer Aggregation Network-You Only Look Once version 8 nano), for real-time sugarcane stem health classification (Fit vs. Unfit). The proposed model achieves ultra-low computational complexity of only 7.2 GFLOPs, 2.2 million parameters, and 4.6 MB model size (in PyTorch)—significantly lower than prior YOLOv8n and all other nano-level YOLO versions—while still providing state-of-the-art performance of 90.1% mAP@50, 64.2% mAP@95, and 14.33 ms average inference time. The model is trained on a self-curated dataset of 3839 images collected under various natural farm conditions in multiple Indian regions. We conduct extensive benchmarking with major YOLO versions, demonstrating our proposed model’s superior efficiency and competitive accuracy. Furthermore, we empirically validate the model’s real-world deployability across multiple edge AI platforms (Jetson Nano, Raspberry Pi 4B) and inference formats (PyTorch, TorchScript, ONNX, TensorRT, MNN and NCNN). The proposed model offers a real-time framework for automated sugarcane stem health detection, providing a crucial step toward intelligent, edge-enabled agricultural automation solutions for better seed quality and higher crop productivity. Source code: https://github.com/NCU-Pushpendra/project-code/blob/main/real-time-sugarcane-detection.ipynb.