Research On Crop Disease Detection Based On Image Recognition And Deep Learning

China is a big agricultural country,crop disease problems seriously affect the national life and national economy,so the rapid and accurate identification method of crop disease is particularly important.At present,crop diseases are recognized by Convolutional neural network(CNN),but the high performance Convolutional recognition model is too large,requires too much hardware equipment and training data,and the Convolutional neural network has insufficient global feature extraction ability.In view of the above problems,the following studies are carried out:(1)Proposed a maize disease recognition method using asymmetric convolution and attention mechanism to optimize residual network.Based on the residual network(Res Net50),the original residual block structure is retained,asymmetric convolution and channel attention mechanism are introduced to reduce the number of parameters of the model,and transfer learning method is adopted to further improve the performance of the model.The experimental results show that the residual network combined with asymmetric convolution and channel attention mechanism can effectively reduce the number of model parameters,shorten the model training time,and the recognition accuracy of corn disease can reach 97.25%.(2)Proposed a tomato disease image recognition method based on Knowledge Distillation and Generative Adversarial Networks(GAN).The problem of insufficient data in deep learning is alleviated by generative adversarial network,and the model is compressed by knowledge distillation.A large-scale and high-precision teacher network was established to guide small students’ network learning target features,thus improving the target recognition characteristics of the small network.Experimental results show that after knowledge distillation,the accuracy of the student model is increased by 2.67%,the number of parameters is reduced by 88% compared with the teacher network,and the single training time is doubled.(3)Vision Transformer method combining Wasserstein distance GAN for apple disease recognition was proposed.Visual converter does not rely on convolution layer for feature extraction,but use,is long attention mechanism as the main unit,the plant disease image is divided into small pieces,then through communication and embedded converts it to a sequence,through the position encoding and visual converter feature encoding,depend on to find out the internal characteristics of the image,the generated sequence will be sent to multiple long attention layer.The experimental results show that the visual converter can still achieve a high accuracy of apple disease recognition without convolutional neural network.The above research starts from the perspective of deep learning model compression and model performance,providing a new method for deep learning in crop disease identification.