Shi Hang, Liu Changxi, Wu Miao, Zhang Hui, Song Hang, Sun Hao, Li Yufei, Hu Jun
College of Engineering, Heilongjiang Bayi Agricultural University, Daqing, China.
Heilongjiang Province Conservation Tillage Engineering Technology Research Center, Daqing, China.
Front Plant Sci. 2025 Apr 3;16:1558378. doi: 10.3389/fpls.2025.1558378. eCollection 2025.
Accurate application of pesticides at the seedling stage is the key to effective control of Chinese cabbage pests and diseases, which necessitates rapid and accurate detection of the seedlings. However, the similarity between the characteristics of Chinese cabbage seedlings and some weeds is a great challenge for accurate detection.
This study introduces an enhanced detection method for Chinese cabbage seedlings, employing a modified version of YOLO11n, termed YOLO11-CGB. The YOLO11n framework has been augmented by integrating a Convolutional Attention Module (CBAM) into its backbone network. This module focuses on the distinctive features of Chinese cabbage seedlings. Additionally, a simplified Bidirectional Feature Pyramid Network (BiFPN) is incorporated into the neck network to bolster feature fusion efficiency. This synergy between CBAM and BiFPN markedly elevates the model's accuracy in identifying Chinese cabbage seedlings, particularly for distant subjects in wide-angle imagery. To mitigate the increased computational load from these enhancements, the network's convolution module has been replaced with a more efficient GhostConv. This change, in conjunction with the simplified neck network, effectively reduces the model's size and computational requirements. The model's outputs are visualized using a heat map, and an Average Temperature Weight (ATW) metric is introduced to quantify the heat map's effectiveness.
Comparative analysis reveals that YOLO11-CGB outperforms established object detection models like Faster R-CNN, YOLOv4, YOLOv5, YOLOv8 and the original YOLO11 in detecting Chinese cabbage seedlings across varied heights, angles, and complex settings. The model achieves precision, recall, and mean Average Precision of 94.7%, 93.0%, and 97.0%, respectively, significantly reducing false negatives and false positives. With a file size of 3.2 MB, 4.1 GFLOPs, and a frame rate of 143 FPS, YOLO11-CGB model is designed to meet the operational demands of edge devices, offering a robust solution for precision spraying technology in agriculture.
在苗期准确施用农药是有效防治大白菜病虫害的关键，这就需要对幼苗进行快速准确的检测。然而，大白菜幼苗与一些杂草的特征相似性给准确检测带来了巨大挑战。
本研究介绍了一种针对大白菜幼苗的增强检测方法，采用了YOLO11n的改进版本，称为YOLO11-CGB。通过将卷积注意力模块（CBAM）集成到其主干网络中，对YOLO11n框架进行了增强。该模块专注于大白菜幼苗的独特特征。此外，一个简化的双向特征金字塔网络（BiFPN）被纳入颈部网络，以提高特征融合效率。CBAM和BiFPN之间的这种协同作用显著提高了模型识别大白菜幼苗的准确性，特别是对于广角图像中的远距离对象。为了减轻这些增强带来的计算负担增加，网络的卷积模块已被更高效的GhostConv取代。这一变化与简化的颈部网络相结合，有效地减小了模型的大小和计算需求。使用热图对模型的输出进行可视化，并引入平均温度权重（ATW）指标来量化热图的有效性。
对比分析表明，YOLO11-CGB在检测不同高度、角度和复杂环境下的大白菜幼苗方面优于Faster R-CNN、YOLOv4、YOLOv5、YOLOv8等已有的目标检测模型以及原始的YOLO11。该模型的精度、召回率和平均精度分别达到94.7%、93.0%和97.0%，显著减少了误报和漏报。YOLO11-CGB模型的文件大小为3.2 MB，4.1 GFLOPs，帧率为143 FPS，旨在满足边缘设备的运行需求，为农业精准喷洒技术提供了强大的解决方案。
