Research On Crop Diseases Based On Regional Convolutional Neural Network

China is a large agricultural country,and agriculture is the basic industry supporting the country's economic development.Therefore,it is the top priority to increase crop output and agricultural production level.However,due to natural environment and natural attributes of crops,crop diseases occur frequently,which seriously affects the quality and yield of crops.Therefore,timely and accurate diagnosis of crop diseases is of great significance for agricultural development in China.At present,the main way of crop disease diagnosis is still the method of artificial identification.Due to the lack of professional knowledge of most agricultural producers,it is easy to misjudge the disease types and use the wrong medicine,which not only has no control effect,but also causes soil pollution.In addition,China's agricultural production is scattered,but the number of agricultural experts is small,professionals cannot guarantee that they can provide timely technical support to agricultural producers,which delays the best opportunity to control diseases and causes irreversible economic losses.In recent years,with the development of computer vision technology and image processing technology,it is possible to diagnose crop diseases through crop disease images.Through the detection of disease sites(fruits and leaves,etc.)in crop disease images,the disease types can be determined.In this paper,the regional neural network Faster R-CNN was used to detect the crop disease images,identify the disease types and accurately locate the location of crop diseases.In addition,the regional convolutional neural network Mask R-CNN was used to segment the targets in the image of crop diseases,which could not only identify the disease types,but also accurately segment the contour of crop disease spots.In view of the lack of crop diseases dataset,a new crop disease images dataset FLD4840 was developed,including 17 kinds of fruit and leaf diseases of 6 common crops.This dataset can be directly applied to the training and detection of regional convolutional neural network model.All experiments in this paper are based on this dataset.When detecting the disease images with Faster R-CNN,in order to improve the accuracy of the model,the idea of residual network was introduced,and the residual network models ResNet-50 and ResNet-101 were used to extract the image features,respectively.The experiments showed that the detection accuracy of Faster R-CNN based on the residual network was higher than that of the original model.In order to meet the requirement of real-time detection in practical application,the depthwise separable convolution was introduced to replace the standard convolution operation in the original model,which greatly shortened the training time,and the recognition accuracy could meet the demand of practical application.When using Mask R-CNN to segment the target in the disease image,the recognition accuracy of the original model is already very good,but the problem of large computation and long training and testing time exists.In order to solve the above problems,the depthwise separable convolution is also introduced to replace the standard convolution operation in the original model,and a lightweight target segmentation network is established,which not only guarantees the accuracy of the model,but also shortens the training time of the model.