基于卷积神经网络的棉花作物行检测

Row Detection of Cotton Crops Based on Convolution Neural Network

针对传统作物行分割算法难以对覆膜回收期棉花作物行进行分割的问题,提出基于卷积神经网络的作物行检测方法。首先使用卷积神经网络U-Net对棉花作物行进行分割,然后从分割结果中定位得到作物行中心点并进行分类,最后基于随机抽样选取最优拟合直线作为作物行中心线。实验结果表明:该算法对多种光照条件具有良好的鲁棒性,棉花作物行检测准确率达到92.68%,对分辨率为640×480图像的处理时间小于70 ms,能够满足实用要求。

To solve the problem that traditional crop row segmentation methods are difficult to segment cotton crop rows during the recovery period of remnant membrane recycling, a cotton crop row detection algorithm based on convolution neural network is proposed. Firstly, convolution neural network U-Net is used to segment cotton crop rows. Then the center points of crop rows are located and classified from the segment result. Finally, an optimal fitting straight line is selected as the center line of crop rows based on random sampling. The experimental results show that the algorithm has good robustness to a variety of light conditions. The detection accuracy of cotton crop line reaches 92.68%, and the processing time of a 640×480 image is less than 70 ms, which can meet the practical requirements.