Research On Identification Of Multiple Crop Diseases Based On Convolutional Neural Network

Crop diseases are one of the main causes of crop yield reduction and directly affect the development of agricultural economy,so it is crucial to carry out accurate and efficient control of crop diseases.Traditional disease identification methods mainly rely on manual experience for judgment,which is highly subjective and prone to problems such as misdiagnosis and delay in the best time for treatment.With the rapid development of computer vision technology,traditional machine learning and deep learning techniques are gradually applied to the field of crop disease identification.However,traditional machine learning methods rely on manual extraction of features,which also have the problem of relying on subjective experience.Compared with manually selected features,deep learning can automatically learn features from sample data,avoiding a certain degree of subjectivity,and convolutional neural network is one of the algorithms in deep learning that can achieve higher recognition accuracy.Based on this,this paper takes crop diseases as the research object,and studies various crop disease recognition methods based on convolutional neural networks,and proposes two models,MSA-ResNet,a multi-scale residual network,and LMSA-ResNet,a light-weight residual network.The main contents are as follows:(1)Crop disease recognition based on MSA-ResNet,a multi-scale residual network.The crop disease spots have various symptoms and are in different locations in the leaves.To address the problem of weak feature extraction ability of single-scale convolution kernel in convolutional neural networks,a crop disease recognition method based on multi-scale residual network MSAResNet is proposed,which uses multi-scale convolution to extract multi-scale features of the disease,and introduces a convolutional block attention module to make the network model more focused on the details of the disease effective features and the location information of leaf diseases,which together constitute the multiscale attentional residual blocks to construct the MSA-ResNet model.Experiments on the AI Challenger 2018 dataset show that the MSA-ResNet recognition accuracy is higher,reaching 89.64%,with better performance compared to network models such as VGG16,ResNet50,and Inception V3.2)Crop disease identification based on Lightweight Residual Network LMSA-ResNet.The crop disease dataset with actual scenarios is used,which has mostly complex background and small sample size,and the problem of small sample size is effectively alleviated by using data enhancement methods and model migration learning strategies.To address the problem of large number of parameters in the current network model,we propose a combination of convolutional decomposition and parametric-free attention module based on MSA-ResNet to improve the multi-scale attention residual block with light weight.Firstly,the large-size convolution kernel in the multi-scale residual block is replaced by a stack of small-size convolution kernels,and part of the standard convolution is replaced by a depth-separable convolution,secondly,the symmetric structure of the convolution layer is improved to an asymmetric structure in the deep layer of the network,which reduces the number of parameters on the one hand,deepens the network depth on the other hand,and improves the nonlinear expression capability of the model,and finally,the parametrization-free attention module is introduced to solve the problem of the convolution block Finally,the attention module is introduced to solve the problem of increasing the number of parameters caused by the convolutional block,and improve the recognition performance of the network.Experiments on the Plant Doc dataset show that the recognition accuracy of LMSA-ResNet proposed in this paper is higher than that of Mobile Net V2 and other models,reaching 64.86%,the number of parameters is effectively reduced to 6.22 M,the memory occupied by the model is relatively small,with a size of 24.6MB,and the recognition speed of a single image is the fastest,at 0.745 s,with a comprehensive performance The performance is relatively good,which is suitable for crop disease identification applications in practical environment and has certain practical value.