Research On Improved Convolutional Neural Network And Its Application In Image Recognition Of Crop Leaf Diseases

In recent years,as the conception of green agriculture was introduced,the prevention and control of crop leaf diseases have received extensive attention.At present,some progress has been made in applying computer vision technology to the field of crop leaf diseases identification.However,traditional computer vision-based diseases recognition methods rely on artificial feature selection,which is difficult to reflect the characteristics of crop leaf diseases comprehensively,and the accuracy still needs to be improved.Convolutional neural network?CNN?is the deep learning algorithm,which can effectively avoid the complex feature selection process of traditional computer vision technology and achieve higher recognition accuracy than traditional algorithms.However,existing leaf diseases recognition models based on deep learning have some problems such as large parameters and so on.In view of this,the pictures of crop leaf diseases under simple background and actual complex background were taken as the research object in this paper,and the traditional CNN model was improved.The details are as follows.?1?To achieve rapid and accurate identification of crop leaf diseases under a simple background,the improved CNN model was proposed.Aiming at the problems of large amount of parameters,calculations and slow convergence speed of CNN,the CNN was improved by combining depth separable convolution and global average pooling to reduce the number of parameters and the amount of calculation;To accelerate the convergence of the model,residual connection and batch normalization were used in the proposed DW-ResNet model.The results showed that the convergence speed of the DW-ResNet model was faster,and the residual connection could improve the accuracy and convergence speed of the model simultaneously.Compared with the crop leaf diseases identification method based on fine-tuning VGG model and traditional machine learning,the proposed DW-ResNet model could achieve higher crop leaf diseases identification accuracy.The test accuracy on the crop leaf diseases dataset and diseases severity dataset could reach 98.59%and 89.16%,respectively.Moreover,the inference time of the DW-ResNet model was only 21 ms.?2?Under a small crop leaf diseases dataset,using deep learning methods to accurately identify crop leaf diseases that in a complex background is difficult.So the transfer learning strategy was introduced in this paper to realize the effective training of the model under small dataset.The pre-trained model of generalized recognition has a large number of parameters and weak ability to extract fine-grained feature.To solve the problem above,the lightweight CNN model was pre-trained with a simple background crop leaf diseases dataset.Aiming at the problem that imbalanced distribution of classes in the field crop leaf diseases dataset,FL_multiloss function was proposed to enhance the learning ability of the model for a few sample categories and difficult samples.The experiment results revealed that theFL_multiloss function proposed in this paper had better performance when dealing with unbalanced dataset.Compared with the method of fine-tuning MobileNet-v2,the DW-ResNet-FL algorithm usingFL_multiloss function had a stronger robustness and anti-interference ability,and its F1-score could still reach 90.60%on the noise-added test dataset.In addition,this model had advantages in terms of occupied memory size,and the occupied memory of DW-ResNet-FL was only 2.1 MB.In summary,the crop leaf diseases identification method based on the improved lightweight convolutional neural network can realize high-precision and rapid identification of crop leaf diseases under the complex background in field.The research results can provide technical support for the accurate and robust identification of crop leaf diseases,and reference technology for the development of crop leaf diseases intelligent identification software,which has certain theoretical research significance and practical application value.