基于YOLOv4的稻田杂草目标检测算法

Weed target detection algorithm in paddy field based on YOLOv4

针对当前精准农业中智能除草设备工作时需要精确快速识别稻田杂草的问题,提出了一种基于YOLOv4算法的稻田杂草检测方法。该方法参照PASCAL VOC数据集格式,构建了稻田杂草目标检测数据集,用深度可分离卷积代替原有的标准卷积,并将逆残差组件(Inverted Residual Unit,IRU)代替CSP-Darknet中的残差组件(Residual Unit,RU),使模型减少参数数量,提高了检测速度。此外,将用K-means算法聚类得到的边界框尺寸应用到各尺度网络层,在路径聚合网络(Path Aggregation Network,PANet)的自适应特征池输出后添加生成对抗网络(Generative Adversarial Network,GAN)噪声层,提高了模型检测精度。将改进的模型在GPU服务器上进行算法训练,并与原始YOLOv4算法进行模型性能的试验对比。结果表明:改进的算法在测试集上的平均精度均值(mean Average Precision,mAP)比原始算法高出4%,达到97%;检测速度提高了12.1帧∕s,达到60.3帧∕s,改进效果明显。该算法具有实时性好、精度高、鲁棒性强的优点,可以更好地实现智能除草设备对稻田杂草的检测,极大节约人力、物力的投入。

Aiming at the problem of accurate and rapid identification of weeds in paddy field during the operation of intelligent weed control equipment in precision agriculture,a method of weed detection in paddy field based on YOLOv4 algorithm is proposed.According to PASCAL VOC data set format,a paddy field weed target detection data set was constructed.Depth separable convolution was used to replace the original standard convolution,and inverted residual unit was used to replace the residual unit in CSP-Darknet to reduce the number of parameters and improve the detection speed of the model.In addition,the boundary box size obtained by Kmeans clustering algorithm was applied to each scale network layer,and the Generative Adversarial Network(GAN)noise layer was added to the adaptive feature pool output of the Path Aggregation Network(PANet)to improve the detection accuracy of the model.The improved model was trained on GPU server and compared its performance with the original YOLOv4 algorithm through experiments.The results showed that the mean Average Precision(mAP)of the improved algorithm was 4%higher than that of original algorithm on the test set,reaching 97%;The detection speed had increased by 12.1 frames∕s,reaching 60.3 frames∕s,with significant improvement effects.It had the advantages of good real-time performance,high accuracy,and strong robustness,which could better achieve the detection of weeds in rice fields by intelligent weed control equipment,greatly saving manpower and material resources.