Research On Crop Disease Image Recognition Based On Deep Learning

China has a wide variety of crops,and crop yield and quality affect economic development and farmers’ income,and the healthy and stable development of agriculture has always been a national goal.Therefore,it is crucial to quickly and accurately identify crop disease types and give accurate control in time.Traditional crop disease image identification mainly relies on a small number of agricultural experts to identify from the color and shape of the disease image,which is laborious,inefficient and subjective,and can easily lead to misjudgment and miss the best treatment time.The existing research methods of crop disease image recognition mainly combine from image background segmentation and machine learning for disease recognition,but the image background segmentation is difficult and the data samples are small,which easily causes inaccurate recognition.Based on the above limitations,this paper proposes a research on crop disease recognition method based on Convolutional Neural Network(CNN)algorithm.In this paper,several crop disease images are studied,and 5 types of corn diseases,5 types of tomato diseases,3 types of potato diseases and 2 types of strawberry diseases are used as experimental research objects.For the problem of large memory and many parameters of convolutional neural network,channel pruning and knowledge distillation algorithms are introduced to solve the problem;for the problem of catastrophic forgetting of convolutional neural network,continuous learning algorithms are introduced to improve it.Finally,a PC interface is developed for automatic crop disease detection.The main work is as follows:(1)A migration learning and model compression method for maize disease identification based on improved VGG16 is proposed.To address the problem of insufficient corn disease image samples,firstly,the dataset is enhanced and expanded;secondly,with the help of migration learning,the network model is trained in the large public dataset Image Net,and in this paper,the VGG16-INCE network model is pre-trained first,i.e.,VGG16 and Inception modules;then the pre-trained network feature parameters are retained to realize the training of common corn disease images,the recognition.Experiments show that the recognition accuracy of disease images using migration learning reaches 93.38% on the Image Net dataset.After migration,the model is compressed by combining channel pruning and knowledge distillation,and the compressed model is then used to recognize corn disease images using migration learning.The experiments show that the recognition accuracy of disease images after compression reaches 92.40%,the accuracy rate decreases by 0.98%,the model size is compressed from 73.90 MB to 9.45 MB,and the parameter amount is reduced by 87.80%.This method can ensure the recognition accuracy in small sample scenarios and further realize the model lightweight.(2)Propose a crop disease identification method based on increasing the generality of the model.It addresses the problem of catastrophic forgetting of the model,which leads to low recognition rate of crop diseases.The method introduces the EWC method in holding learning,which enables the model to retain important parameters when training crops,mitigates the phenomenon of catastrophic forgetting of previous tasks,and can improve the crop recognition accuracy.It is shown through experiments that the recognition accuracy of the network model decreases with the increase of crop disease types when no EWC is added,while the proposed method in this paper can continuously achieve recognition of multiple crop disease images with higher recognition results than the results without the addition of continuous learning,and the effectiveness of the method in this paper is verified through experiments on three crop disease data.