无线电工程

在YOLOv5深度神经网络基础上融合CA注意力机制，设计了田间杂草自动识别模型。对不同天气、不同背景下拍摄的包含农作物的杂草图片进行背景分割和数据增强等预处理后，利用深度卷积神经网络进行特征提取，采用随机梯度下降进行模型训练，并与YOLOv4、SSD、Faster R-CNN等方法进行对比实验，改进后的杂草模型具有更强的分类和检测能力，基本解决杂草与农作物相似度较高的问题。实验结果显示，融合CA注意力机制的YOLOv5的杂草识别模型大小为14.1 MB,识别精度达到0.911,召回率为0.950,F₁为0.88,平均准确率为0.904,单张杂草图像检测平均耗时仅为20 ms,整体性能最好。模型能够准确识别出独立、与农作物交叉和贴近等不同空间位置的长茎尖叶杂草，为作物生长精准管理和田间精准喷药提供技术支持。

An automatic field weed recognition model is designed based on YOLOv5 deep neural network incorporating CA attention mechanism. After pre-processing with the background segmentation and data enhancement for weed images containing crops taken under different weather and backgrounds, feature extraction is performed using deep convolutional neural networks, while the stochastic gradient descent is used for model training. Comparing with YOLOv4, SSD, Faster-RCNN and other methods, the improved weed model has stronger classification and detection ability, and basically solves the problem of high similarity between weeds and crops. Experimental results show that the weed recognition model of YOLOv5 fused with CA attention mechanism has a size of 14.1 MB, with a recognition accuracy of 0.911, a recall rate of 0.950, the F₁ of 0.88, and an average accuracy of 0.904. The average time spent for single weed image detection is only 20 ms, which is the best overall performance. The model can accurately identify long-stemmed, sharp-leaved weeds in different spatial locations such as independent, intersecting with crops, and close to crops, providing technical support for accurate crop growth management and precise spraying in the field.