Scholarlink

<p>This article addresses limitations in YOLO-based sugarcane bud detection by balancing model size, speed, and robustness in complex agricultural settings. This article introduces EfficientConv, a lightweight, drop-in convolutional block that is 9% less computationally expensive than standard Conv layers used in YOLO models. By combining Ghost and Shuffle Convolutions with Squeeze-and-Excitation and channel-wise attention, it reduces computation without compromising accuracy. EfficientConv is version-independent and integrates seamlessly into the YOLO architecture. We evaluate YOLOv8n to YOLOv12n models with and without EfficientConv on a proprietary sugarcane bud dataset and test generalization on the VisDrone benchmark. Real-time performance is benchmarked on edge devices such as Jetson-Nano and Raspberry Pi-4B. The performance gains varied with the number of Conv blocks in each YOLO variant. Since YOLOv8n, YOLOv11n, and YOLOv12n have a large number of Conv blocks (six Conv blocks in each variant) located at different stages, they benefited the most from EfficientConv integration, making both the detection and the computing efficacy significantly improved. On the contrary, YOLOv9n and YOLOv10n have only one and two Conv blocks, respectively, and exhibit comparatively modest improvements. YOLOv8n with EfficientConv achieved the best results: 73.5% mAP@50, 49.8% mAP@95, and a 12.4% GFLOPs reduction. YOLOv11n and YOLOv12n also used 9.1% and 7.6% fewer GFLOPs while maintaining or improving accuracy. This article advances precision agriculture by introducing a fast, reliable, and efficient sugarcane bud detection. The proposed EfficientConv block reduces GFLOPs and model size while preserving or enhancing accuracy, supporting automation in sugarcane farming. The source code is available at: <a href="https://github.com/Pushpendra-kiet/codebase.git">source</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 Bud Detection Using Efficient Lightweight Convolutional Blocks

  • 期刊论文

  • 发表日期：

    2025年12月5日

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

摘要

This article addresses limitations in YOLO-based sugarcane bud detection by balancing model size, speed, and robustness in complex agricultural settings. This article introduces EfficientConv, a lightweight, drop-in convolutional block that is 9% less computationally expensive than standard Conv layers used in YOLO models. By combining Ghost and Shuffle Convolutions with Squeeze-and-Excitation and channel-wise attention, it reduces computation without compromising accuracy. EfficientConv is version-independent and integrates seamlessly into the YOLO architecture. We evaluate YOLOv8n to YOLOv12n models with and without EfficientConv on a proprietary sugarcane bud dataset and test generalization on the VisDrone benchmark. Real-time performance is benchmarked on edge devices such as Jetson-Nano and Raspberry Pi-4B. The performance gains varied with the number of Conv blocks in each YOLO variant. Since YOLOv8n, YOLOv11n, and YOLOv12n have a large number of Conv blocks (six Conv blocks in each variant) located at different stages, they benefited the most from EfficientConv integration, making both the detection and the computing efficacy significantly improved. On the contrary, YOLOv9n and YOLOv10n have only one and two Conv blocks, respectively, and exhibit comparatively modest improvements. YOLOv8n with EfficientConv achieved the best results: 73.5% mAP@50, 49.8% mAP@95, and a 12.4% GFLOPs reduction. YOLOv11n and YOLOv12n also used 9.1% and 7.6% fewer GFLOPs while maintaining or improving accuracy. This article advances precision agriculture by introducing a fast, reliable, and efficient sugarcane bud detection. The proposed EfficientConv block reduces GFLOPs and model size while preserving or enhancing accuracy, supporting automation in sugarcane farming. The source code is available at: source