Research On Cucumber Disease Recognition Method Based On Deep Learning

Cucumber is an important vegetable variety in the common people’s vegetable basket,which has a wide planting range and a large yield in China.Cucumber disease control is a key link in the effective improvement of cucumber yield,and accurate disease diagnosis is also an indispensable part of the control process.The disease symptoms of cucumber are mainly manifested in the leaf information.The physiological structure and morphological characteristics of the leaves will change when the plant is infected,and the degree of disease information is similar,which makes the manual identification face some difficulties.Machine vision research based on deep learning has made great breakthroughs in recent years,and the detection application of disease spots and disease has become more and more mature.It is of great significance to study disease diagnosis methods by using computer vision and deep learning related technology,which is of great significance to the prevention and control of crop diseases and the quality improvement of agricultural products..This paper aims to obtain the diagnosis method of cucumber disease with high accuracy through the integration of deep learning model.The specific work content is as follows:1.Analysis of deep learning images in the field of several computer vision tasks,including target detection,image classification and segmentation in the information provided in disease diagnosis,considering feature fusion,model accuracy and other factors for the selection of network framework,the classification,detection,segmentation of three visual model integration,discuss the feasibility of composite diagnosis model.2.Through greenhouse field shooting,the original cucumber leaf data set,divided into disease data set and disease data set using against generating network amplification,lableme instance segmentation annotation,disease data set using simple data enhancement method,including scale transformation,brightness transformation,affine transformation.3.Mask-Rcnn network with both object detection and instance segmentation functions is selected as deep learning model,Res Net + FPN as feature extraction network;add image classification branches,common feature extraction network,image classification information and segmentation mask are output as composite diagnostic information.4.Recall,accuracy,m IOU are added as the evaluation criteria;Dropout and regularization are added to solve the numerical stability problem,and the non-maximum suppression algorithm and feature extraction layer are improved according to the disease diagnosis requirements to optimize the segmentation and classification effect.The optimized average accuracy of disease classification reached95.5%,which increased by 2.4 percentage points from the original model,and the m IOU of disease spot segmentation reached 93.1%,which increased by 1.2percentage points compared with the original model.