Research On Recognition Of Typical Leaf Diseases Based On Images And Spectral Information

With the rapid and continuous development of information technology,standardization,digitization,no damage,real-time monitoring and identification of plant diseases have become the development trend of plant disease diagnosis.In recent years,with the continuous expansion of the field cropping area,the field crops have been gradually affected by various diseases and insect pests,which seriously threatens the quality and output of the field crops,and the economic income of farmers has also been affected to varying degrees.A large number of studies have shown that most of the parts affected by plant diseases and insect pests are the surface of the plant leaves,which affects the normal growth of the leaves,and different diseases have different characteristics on the leaves in terms of contour,color,texture and other characteristics.Therefore,how to efficiently predict and diagnose crop leaf diseases is one of the problems that need to be solved urgently.However,most of the current crop disease diagnosis still relies on the observation and work experience of professionals to make judgments.The efficiency is not high and the accuracy and timeliness of disease diagnosis cannot be achieved.In view of the above problems,this article comprehensively applies knowledge in many fields such as image processing technology,hyperspectral technology,spectral analysis technology,etc.to carry out research on rapid and non-destructive recognition methods for major diseases of soybean,corn and ginseng leaves,and on this basis Designed and established a crop leaf disease recognition system.The main research contents of this article are as follows:(1)The data volume of the disease image is expanded and preprocessed.According to the characteristics of the diseased leaf image data set,the original image is expanded by geometric transformation and intensity transformation,and the image is preprocessed,including image size scaling,image enhancement and image normalization.(2)The convolutional neural network is used to classify and recognize the diseased leaf images.Two convolutional neural network models,LeNet and AlexNet,were used to train the data set,and their Accuracy rates were 84.59% and 91.04%,and the loss function values were 0.1017 and 0.0637,respectively.Finally,the AlexNet model was selected.And the spatial pyramid pooling layer is introduced to improve the existing AlexNet model,so that the improved model recognition rate reaches93.91%,and the loss function Loss value drops to 0.0331.(3)Using spectrum detection technology to classify and identify diseased leaves.Using stepwise discriminant method,515 nm,516nm,521 nm,522nm,523 nm,524nm,528 nm,594nm,598 nm,667nm,668 nm,738nm,803 nm,843nm,854 nm,871nm,880 nm,882nm were selected from the spectral curves.18 feature wavelengths form a feature space,and a linear distance discriminant model is established,and the recognition accuracy rates of training samples and test samples are 100% and 94.2%,respectively.(4)The preliminary design completed the recognition system of crop disease leaf.In the integrated development environment of Windows system and PyCharm,the mixed programming technology of Python and Java was used to realize the development of crop disease leaf recognition system,and the system was tested.Test results show that this system can be used in practice.In this paper,the combination of image and spectral information is feasible for early detection and rapid identification of crop leaf diseases.It provides examples for the application of image processing technology and spectral detection technology in agriculture.Disease recognition has certain reference and practical significance.