基于深度学习的“一膜两年用”玉米全膜双垄沟种床特征识别方法研究

Feature recognition method of corn full-film double-ridge furrow with“one film used for two years”based on deep learning

为更加准确掌握“一膜两年用”玉米全膜双垄沟覆膜种床特征,提高其智能化生产水平,采用基于深度学习的目标检测网络模型开展“一膜两年用”玉米全膜双垄沟种床特征识别方法的研究,对种床结构中的残膜、根茬和覆土带进行检测识别。结果表明:优化后的YOLOx网络模型的效果相关值(mAP)为90.76%,检测结果优于其他网络模型。为探究无人机拍摄图像对目标检测结果的影响,采用三因素三水平BOX-Behnken试验设计方法,建立无人机飞行高度、无人机飞行角度和无人机飞行速度与评价网络模型mAP值的数学模型,寻求无人机拍摄的最优参数组合并进行试验验证,当无人机飞行高度为0.9 m、飞行角度100°、飞行速度2.2 m·s^(-1)时,评价网络模型效果相关的响应值达到最优。经验证,无人机拍摄参数优化后,网络模型的mAP值为92.67%,检测效果优于其他组模型。

To more accurately grasp the characteristics of corn full-film double-ridge furrow for“one film used for two years”and improve its intelligent level,based on deep learning of target detection network model,feature recognition method of corn full-film double-ridge furrow for“one film used for two years”was employed,and the residual membrane,root stubble and overburden zone in the seed bed structure were detected and identified.The results showed that the mean average precision(mAP)of network model of YOLOx was 90.76%,which was better than other network models.In addition,to explore the impact of image acquisition of drone on target detection results,a mathematical model related to mAP was established based on the three factor and three level of BOX Behnken experimental design method,which included three parameters:flight altitude,flight angle and flight speed of drone.The optimal parameter combination of drone for image acquisition was found through experiments and verified.The results showed that,the mAP was maximum when the drone flied at a height of 0.9 m,a flight angle of 100°,and a speed of 2.2 m·s^(-1).After optimizing the shooting parameters of drone,the mAP of the network model was 92.67%,which was better than other groups.