Authors

Abstract

Addressing issues such as high soybean seedling detection omission rates and inaccurate target recognition during mechanical weeding operations in soybean fields, which lead to low weeding efficiency, this paper proposes a lightweight convolutional model based on an improved YOLO11 model. Deployed on an intelligent mechanical soybean weeding robot, it utilizes precisely identified soybean seedling coordinates to perform mechanical weeding operations, thereby enhancing weeding efficiency.Building upon the original YOLO11 architecture, this model replaces standard convolutional blocks with Deep Separable Convolution (DWconv) modules. It performs channel pruning on the C3K2 lightweight convolutional module and employs Point-Shuffle operations for channel mixing to enhance feature map information flow, thereby improving edge feature recognition for small targets.The introduction of an Efficient Channel Attention (ECA) mechanism increases channel selectivity for large target features, enhancing sensitivity to critical semantic information. The original loss function is optimized by incorporating an improved bounding box loss function (SIOU), accelerating model convergence and strengthening generalization capabilities.The improved YOLO11 model achieved a 2.0 percentage point increase in mAP50% on the self-built soybean dataset compared to the original YOLO11, reaching 94%. Model parameters and floating-point operations were reduced from 2.59MB and 6.4×10⁶ to 1.97MB and 5.0×10⁶ respectively, representing decreases of 23.9% and 21.9%. This achieves synergistic optimization of model lightweighting and computational efficiency while maintaining detection accuracy.

Abstract in Chinese

针对大豆田间机械除草作业时识别大豆苗漏检率高，识别目标不准确等导致除草效率低等问题，本文提出了一种基于改进YOLO11模型的轻量化卷积模型，部署在智能机械式大豆除草机器人上，利用识精准别到的大豆苗坐标来进行机械除草作业以提高除草效率.该模型在原YOLO11网络架构基础上，使用深度可分离卷积模块DWconv替代普通卷积块，对C3K2轻量级卷积模块进行通道裁剪，使用Point-Shuffle操作进行通道混洗提高特征图间的信息流动，提高对小目标的边缘特征识别效果.引入高效通道注意力机制(ECA)，增大对大目标特征的通道选择性，提高对关键语义信息的敏感度.对原损失函数进行优化，引入改进的边界框损失函数（SIOU），提高模型收敛速度，增强模型泛化性.改进后的YOLO11模型,相较于原YOLO11在自建大豆数据集上mAP50%提高了2.0个百分点，达到了94%.模型参数量、浮点计算量由2.59MB、6.4×106降低至1.97MB、5.0×106同比减少了23.9%和21.9%，在保证检测精度的同时，实现了模型轻量化与计算效率的协同优化.