摘要
针对西南地区多作物套种场景中杂草多样、作物-杂草形态相似等挑战,提出基于YOLOv8n-seg改进的CSM-YOLO模型。通过CA注意力机制增强几何特征,SCConv卷积抑制背景噪声和Matrix NMS加速推理优化模型,构建“作物-环境-杂草”耦合图像数据集进行验证。结果表明:杂草识别精度较基线模型提升0.1%~1.3%,推理速度达197.36 FPS。模型精确率、mAP@0.5、召回率分别为92.9%、94.0%、90.7%,均高于主流模型。可视化效果优于主流模型,尤其在杂草小目标检测任务中表现最优,热力图显示对作物与杂草区域具有高度选择性关注。可为大豆玉米间套种植的杂草识别提供高效的技术支持。
To address the challenges of diverse weeds and high morphological similarity between crops and weeds in intercropping systems in southwestern China,an improved CSM-YOLO model based on YOLOv8n-seg was proposed.The model incorporates a CA attention mechanism to enhance geometric feature extraction,SCConv convolution to suppress background noise,and Matrix NMS to accelerate inference.A“crop-environment-weed”coupled image dataset was constructed for validation.The results showed that the model achieves a 0.1%~1.3%improvement in weed recognition accuracy compared to the baseline model,with an inference speed of 197.36 FPS.The model's precision,mAP@0.5,and recall rate reached 92.9%,94.0%,and 90.7%,respectively,outperforming mainstream models.Visualization results further confirmed its superior performance,particularly in small-target weed detection tasks,with heatmaps indicating highly selective attention to crop and weed regions.This study provides efficient technical support for weed recognition in soybean-corn intercropping systems.
作者
朱惠斌
方圆
白丽珍
王明鹏
李仕
李镕东
ZHU Huibin;FANG Yuan;BAI Lizhen;WANG Mingpeng;LI Shi;LI Rongdong(Faculty of Modern Agricultural Engineering,Kunming University of Science and Technology,Kunming,Yunnan 650500,China)
出处
《干旱地区农业研究》
北大核心
2025年第6期248-258,共11页
Agricultural Research in the Arid Areas
基金
云南省自然科学基金项目(202401AS070115)
国家自然科学基金项目(52265033,51865022)。
