Grape Leaf Segmentation Based On Deep Learning

The growth of grapes under natural conditions,due to the complex weather changes and uneven illumination,will lead to the image foreground and background of grape leaf contrast is not strong,blade image texture blurred,large color difference and other defects.At the same time,in the above conditions,some of the leaves will appear disease,so for how to divide the leaves of the disease and normal,the use of traditional image segmentation algorithm can not solve such problems.In this paper,two kinds of deep learning neural network algorithms are used as the core,the foreground/background segmentation and case segmentation of grape leaf image are realized respectively,and the image segmentation effect is better for various imaging conditions and backgrounds.In order to solve the foreground/background automatic segmentation of normal and diseased grape leaf images under different illumination and complex background,an automatic segmentation algorithm of grape leaf image with full convolution network(FCN)is used.In the structure,the traditional convolution neural network(CNN)is converted into three convolution layers with three full connections.Through Multi-layer convolution,the characters of the imported grape leaf image are extracted,the feature information is filtered through the pool layer,and the feature size is reduced so as to reduce the network parameters.Then,the characteristics of the original dimension are sorted by pixel by deconvolution,and the feature is recovered from the high dimension and small dimension to the original size of the image.Determines whether the label for each pixel position in the original image is a background or a foreground.But only after the sample processing of the segmented image will be more coarse,therefore,by jumping structure,the rough original image is integrated with the local information and the whole information,and the segmentation result is refined,and the results show that the algorithm achieves good segmentation results for both normal and diseased blades.For the effect of no occlusion and occlusion of normal grape blades in sunny,cloudy and shaded conditions,through the MCC evaluation Standard,The average accuracy is 0.9018 and 0.8059,0.8629 and 0.8036,0.8589 and 0.8016,respectively,13.74%,10.47%,18.08%,20.1%,30.55% and 29.73% respectively,compared with the joint segmentation algorithm.For the single blade and multiple blades with disease,the accuracy of the MCC segmentation is 0.8214 and 0.7472,the average is 0.7843,and the segmentation accuracy in three cases is increased by 5.14%,4.39% and 4.76% respectively,compared with the detection based segmentation algorithm.The Division accuracy of ACC can reach 0.9481,0.8774,average 0.9128,and also improve 3.99%,3.18% and 3.59% respectively compared with the contrast algorithm.In this paper,a case segmentation of grape leaf image is realized by using mask r-cnn,first the candidate region(ROI)is generated by RPN(regional convolution network),then the whole feature of the image is extracted by using the convolution layer of fast r-cnn,and then the feature map of the image is obtained,and then the ROI Align further to the feature map pixel correction,the correction of each ROI prediction,get its category and bounding box at the same time each ROI using a well-designed FCN framework to predict the ROI region of each pixel-belonging category,through the IOU(The evaluation function is used to classify the candidate area background and foreground,and to give the label to the 101-layer depth residual network res,so as to realize the result of detection,recognition,segmentation and one case segmentation.For the effect of no occlusion and occlusion of normal grape blades in sunny,cloudy and shaded conditions,the image foreground/background segmentation accuracy of the whole blade is 0.9347 and 0.9002,0.9162 and 0.8951 by analyzing the foreground/background of the whole leaf image and the MCC evaluation standard of each blade.0.9289 and 0.9104,the average value is 0.9143,each grape image of the blade average segmentation accuracy of 0.9414 and 0.9232,0.9258 and 0.9008,0.9356 and 0.9123,the average value of 0.9237,compared to the entire blade image foreground The precision of the segmentation of the background,the average accuracy of each blade in the blade image is higher than that of the former 0.94%.For the single blade and multiple blades with disease,the image foreground of the whole blade is evaluated by the image foreground/background of the whole leaf and the MCC evaluation standard of each blade.The segmentation precision of the background is 0.9237 and 0.8486,the average value is 0.8862,the average value of the blade in each grape image is 0.92374 and 0.8867,the mean value is 0.9052,the average precision is higher than the former 1.9%.Compared with the full convolution network algorithm,the Mask r-cnn is used to evaluate the segmentation accuracy of the normal and diseased blades and different varieties of grape leaves in the mixed-blade image,The accuracy of grape leaf was 0.9513,and the image accuracy of normal grape was 0.9061.For different varieties of leaf images,the average precision of 4 varieties of grape leaves was reached 0.9391,0.8907,by MCC on the whole leaf image foreground/background.0.8873 and 0.9312,the average value is 0.9121;the same grape leaf image carries on the MCC evaluation of each blade,with an average accuracy of 0.9483,0.9092,0.8932 and 0.9432,the average is 0.9235,the accuracy of the two is compared,the latter is higher than the former 1.14%.