基于深度学习的玉米田间杂草识别模型研究

Deep learning-based weed recognition model in the maize field

【目的】针对现有田间杂草识别模型复杂、准确率不高等问题,研究玉米田间杂草识别算法,通过准确识别杂草图像,为提高田间杂草防控效果提供理论与技术支持。【方法】基于深度学习目标检测的方法,选取玉米田间4类常见杂草(早熟禾、藜、刺儿菜、莎草)作为试验数据集,建立了YOLOv3、YOLOv5、SSD目标检测模型,并进行了模型训练。【结果】YOLOv3模型的平均精准率为0.734,平均召回率为0.814,平均F1得分为0.789,mAP值为0.792;YOLOv5模型的平均精准率为0.914、平均召回率为0.967、平均F1得分为0.942、mAP值为0.961;SSD模型的mAP值为0.907。【结论】YOLOv5模型的mAP值为0.961,且其各项指标均优于YOLOv3和SSD目标检测模型。YOLOv5模型更适合用于作物田间精确除草的自动化作业。

【Objective】In response to the complexity and low accuracy of existing field weed recognition models,a corn field weed identification algorithm is studied.By accurately identifying weed images,theoretical and technical support is provided to improve the effectiveness of field weed control.【Method】In this paper,based on deep learning method,four types of common weeds in maize field,bluegrass,chenopodium album,clrsumsetosum and sedge were selected as experimental data sets,and the YOLOv3,YOLOv5 and SSD target detection models were established and trained.【Result】The results showed that the YOLOv3 model achieved precision of 0.734,mean recall of 0.814,mean F1 score of 0.789,and mAP value of 0.972;the YOLOv5 model achieved precision of 0.914,mean recall of 0.967,mean F1 score of 0.942 and mAP of 0.961;the mAP value of the SSD model is 0.907.【Conclusion】The test results show that the mAP value of YOLOv5 model is 0.961,and all of its indexes are better than those of YOLOv3 and SSD target detection model,so YOLOv5 model is more suitable for the automated operation of accurate herbicide spraying in crop field.