Identification Of Crop Leaf Diseases Based On Multi-Scale Augmented Generative Adversarial Network

Plant control includes prevention,immunization,eradication and quarantine.Plant quarantine is the most effective and economical method in the protection,prevention,radical cure and immunization of plants.However,the characteristics of different plant diseases are different,and the prevention and control methods of different diseases are also different,and they have the characteristics of wide transmission range and route,strong infectivity,great harm and short time of epidemic prevention and control.It is difficult to obtain a large number of sample data in time,and the positive and negative data of samples are not balanced.Therefore,the effect of direct identification and quarantine with deep learning method is not good,In this paper,the background of small sample real data is removed by pix2 pix,and the image is binarized by adaptive threshold method.The binarized image and the de background image are made into a pair of image database,which is used as the training data for generating new leaves by pix2 pix,The proportion of healthy leaves and diseases was about 1:1,and the expanded sample set reached ten thousand level.Using transfer learning,vgg16 and RESNET convolution models after image pre training are trained with expanded data.Finally,the final classification model is obtained by fine-tuning the saved network framework with real data.The recognition rate and generalization ability of the model are improved.Finally,the isknet is better in recognition accuracy,although the training time is longer,But it can meet the intelligent classification and detection effect of small sample plant diseases.In this paper,tomato,apple,grape,potato,pepper leaf diseases as the starting point,the main contents are as follows:(1)analysis of the performance of mainstream deep learning and traditional image classification technology in identifying crop diseases;(2)By using pix2 pix method for image preprocessing,the background of crop leaves is removed and used as paired data for the training of confrontation network generator.Finally,the effect of plant leaves generation is better;(3)Vggnet and RESNET deep convolution neural network models are used to train and recognize the plant leaf data,and the results are compared with those of ISKNetwork;(4)The Gan is improved,and a multi-scale enhanced generation countermeasure network Megan is proposed.By adding high-frequency and low-frequency signal processing units and attention mechanism,the extraction of important features such as the texture,size and shape of leaves and disease spots is strengthened,so that the quality of generated leaves is higher,and the purpose of effectively expanding data is achieved;(5)In order to broaden the network and improve the recognition rate,the multi classification model is built by adding the induction module into the sknet model.In this paper,the method is verified by using the leaves of 23 diseases of 5 crops,and the accuracy of disease identification of plant leaves can reach 96.4%.The experimental results show that this method can improve the accuracy of leaf recognition in a small number of samples,and realize the automatic diagnosis of potato,tomato,pepper,grape and apple diseases,with high recognition rate and strong universality.