基于图像处理和深度学习的葡萄叶片钾含量的检测方法

Detection Method of Potassium Content in Grape Leaves Based on Image Processing and Deep Learning

为快速无损地获得红地球葡萄在不同生育期内钾素营养水平,基于图像处理技术和深度学习研究其钾含量检测方法。将试验地划分为A、B两块不同钾素处理区域,在不同生育期内,利用数码相机在自然光下对叶片进行拍照采样,采摘样本经晾晒干燥、磨粉、装袋后送检。运用Photoshop通过直方图程序获取叶片图像的色彩信息R、G、B值,并计算归一化NRI、NGI、NBI及H、S、I均值,实验选取R/(R+B-G)、G/(R+B-G)、B/(R+B-G)构成颜色特征参数,采用Excel对颜色特征值和钾含量进行回归分析。将原始图像经仿射和强度变换后扩展为试验训练数据集,采用改进的YoLoV3-M1卷积神经网络进行模型训练,模型检测准确率达90%以上。试验结果表明,该方法可用于葡萄叶片钾含量的无损检测。

In order to quickly and non-destructively obtain the potassium nutrition level of Red Global Grapes in different growth periods,image processing technology and deep learning were used to study its potassium content detection method.The test site was divided into two different potassium treatment areas,A and B.During different growth periods,digital cameras were used to takes photos and samples of leaves in the natural light.The picked samples were dried,ground,and bagged before being sent for inspection.Using Photoshop to obtain the color information R,G,and B values of the leaf images through the histogram program,and calculate the normalized NRI,NGI and NBI and the average values of H,S and I.For the experiment,R/(R+BG),G/(R+BG)and B/(R+BG)were selected to constitute color characteristic parameters,and Excel was adopted to perform regression analysis of color characteristic values and potassium content.After affine and intensity transformation,the original image was expanded into a test training data set,and the improved YoLoV3-M1 convolutional neural network was used for model training,with the accuracy of model detection being as high as 90%.The test results show that this method can be used for non-destructive testing of grape leaf potassium content.