Research On Deep Learning Architectures For Automatic Classification Of Plant Diseases Based On Leaf Images

The world is facing the utmost challenge of feeding its population with limited and shrinking agricultural resources.One of the ways to overcome this challenge is early,prompt,and accurate recognition of the infested plants.Identification and classification of plant diseases based on visual traits using image processing techniques have evolved by solving various computer vision tasks.Classification of objects based on visual traits is a basic application domain in computer vision.Recent researches have been focusing on the early detection and identification of plant diseases based on visual features which are present due to various abnormalities infested plant parts.These diseased plants pose a potential threat to crop yield and the economy.The application of algorithms to detect abnormalities in plants automatically avoids the need for expert human intervention.Various methods for the classification of plant diseases based on leaf images are presented in this research.Recently,the rapid outbreak of Covid-19 across the world demanded a fast,accurate,and easy detection of the disease.The models used for plant disease detection were further extended to identify and classify human disease,i.e.Covid-19 which was based on X-ray images of the patients.The extension of these models to human studies not only corroborates the previous results of this study but also expands the reach and potential of deep learning algorithms in medical research.The binary classification method of fine-grained leaf images based on a deep convolution neural network is studied.Four classification models,viz: Lenet,Alexnet,VGG16,and VGG19,were developed for diseased plant leaf identification based on a deep convolution neural network.These four models were evaluated and compared.The evaluated data set contains 8 plants and 25283 images.The results show that the deep convolution neural network has a good effect on the binary classification of fine-grained images,and the performance of the deeper model is better than that obtained from shallower models.There is a trade-off in accuracy between computational cost and accuracy.The classification method of crop diseases based on depth separable convolution network was studied.Aiming at the disadvantage that the traditional convolutional neural network needs huge computing resources,a depthwise separable convolutional neural network structure is introduced to reduce the network parameters.Compared to the traditional VGG network,the accuracy of the proposed network matches the classification accuracy of the plant disease dataset,and the parameters are reduced by 29 times while classifying 82161 plant leaf images distributed over 55 classes.To further weigh the relationship between accuracy and parameters,this study proposes two different network structures,Modified Mobilenet,and Reduced MobileNet,focusing on higher accuracy and smaller parameters respectively.These models are suitable for resourceconstrained mobile devices and are helpful for the automatic diagnosis of crop diseases.The classification method of crop diseases based on transfer learning is studied.Aiming at the disadvantage that the traditional deep learning method needs a large number of training data samples,the idea of transfer learning is introduced,that is,reusing the characteristics of the previously learned model on a large-scale data set.In this paper,the pre-trained model on Imagenet is used as the feature extractor,and the model is later fine-tuned on smaller crop disease datasets.The model achieves high accuracy on the premise of a limited number of data samples.In this study,the frozen convolution layer group is used to replace the frozen single convolution layer to reduce the training time.Finally,the results of six common convolutional neural network model structures in crop disease classification by transfer learning are compared,and the optimal classification model is obtained and further tested on other public data sets,which proves that the method is efficient when applied to different data sets.Moreover,the pretrained models were fine-tuned to classify human chest X-ray images to diagnose patients with Covid-19.The influence of background removal on a convolutional neural network is studied.Given the influence on the classification accuracy caused by the background of plants,soil,rocks,and the human body in the leaf image,a method based on edge recognition,morphological segmentation,background removal,and convolution neural network is adopted to improve the accuracy of model classification.Firstly,the method of edge recognition is used to segment the foreground and background area.Secondly,the background is removed by the background removal method to segment the region of interest.Finally,the image classification task is completed through the pre-trained classification network.The experimental results show that the accuracy of this combined method is about12% higher than that of the direct training classification network,and has a faster convergence speed.To sum up,based on the above research work proposed in this paper,the classification task of crop diseases and human diseases is realized,the accuracy of the traditional neural network model is improved,and the parameters and hardware calculation performance requirements required by the classification task model are reduced,and technical support is provided for the automatic identification and diagnosis of crop diseases,and further experimentation of those models on human diseases have been conducted.It has very important theoretical research significance and practical application value.