基于嵌入式YOLO的玉米苗与杂草检测轻量化模型BACD-YOLO研究

Research on Lightweight BACD-YOLO Model for Corn Seedling and Weed Detection Based on Embedded YOLO

摘要:

针对玉米2~5叶除草关键窗口期存在的田间复杂环境、杂草识别精度与检测效率难以兼顾的难题,以东北2~5叶期玉米幼苗及其早期伴生杂草为对象,提出了基于一种改进YOLOv8n的玉米苗及杂草轻量化检测模型BACD-YOLO(BiFPN-Adown-CA-DualConv-YOLO)。该方法采用Adamax优化器增强模型在复杂田间环境下的鲁棒性;引入加权双向特征金字塔网络(Bidirectional feature pyramid network,BiFPN)至特征融合网络作为连接层,改善模型对长势各异杂草的检测效果;采用轻量化下采样模块(Adown)替换网络中常规卷积,降低模型冗余的参数计算量;在SPPF层前及特征融合网络中嵌入坐标注意力机制(Coordinate attention,CA),提高模型对小目标及密集分布杂草的定位能力;采用DualConv轻量化双卷积替换原模型中普通卷积结构,进一步实现模型轻量化及模型对相似特征杂草的检测能力。试验结果表明,改进模型的精确率、召回率及平均精度均值分别为86.6%、86.2%和91.2%,较初始模型分别提高了2.2、1.5、1.6个百分点,且浮点计算量和参数量仅为6.2×109和2.3×106,较初始模型分别降低了23.5%和23.3%。通过主流检测模型对比、公开数据集验证和边缘设备部署应用试验,结果表明,改进模型(BACD-YOLO)凭借较高的检测精度、较强的泛化能力和较优异的轻量化性能适宜边缘设备部署应用,帧率达到19.4f/s,检测正确率为86.6%,能够满足田间实时检测要求。本研究可为玉米苗及杂草精准识别与机器人除草作业提供有效的轻量化解决方案。

Abstract:

Aiming at the problems of complex field environment and difficulty in taking into account the accuracy of weed identification and detection efficiency in the critical window period of 2 ~ 5 leaf weeding of maize, a lightweight detection model BACD-YOLO for maize seedlings and weeds based on improved YOLO v8n was proposed. Using Adamax optimizer to enhance the robustness of model in field environment; the weighted bidirectional feature pyramid network (BiFPN) was introduced to the feature fusion network as the connection layer to improve the detection effect of the model on weeds with different growth. Adopting lightweight down sampling module (Adown) replaced the conventional convolution in the network to reduce the amount of parameter calculation of model redundancy; the coordinate attention (CA) mechanism was embedded in the SPPF layer and the feature fusion network to improve the positioning ability of the model for small targets and densely distributed weeds; DualConv lightweight double convolution was used to replace the ordinary convolution structure in original model to further realize the lightweight of the model and the detection ability of the model for weeds with similar characteristics. The experimental results showed the accuracy, recall and average accuracy of improved model were 86. 6%, 86. 2% and 91. 2%, respectively, which were 2. 2, 1. 5 and 1. 6 percentage points higher than that of the original model, and floating-point calculation and parameter quantity were only 6. 2 × 109 and 2. 3 ×106, which were 23. 5% and 23. 3% lower than that of original model, respectively. According to the verification test results, the improved model was more suitable for edge device deployment application with high detection accuracy, strong generalization ability and excellent lightweight performance. The frame rate was 19. 4 f/ s, and the detection accuracy was 86. 6%, which can meet requirements of field real-time detection. The research result can provide an effective lightweight solution for accurate identification of corn seedlings and weeds and robot weeding.