| As the foundation of China’s national economy,agriculture is one of the most important industries in China.Crop pests and diseases seriously affect the yield and quality of agricultural products,thus affecting the development of the agricultural economy.Therefore,the early realization of rapid and accurate identification of crop pests and diseases is of great significance for later control.Traditional crop pest identification usually relies on visual observation by agricultural experts,a method based on subjective human perception,which can easily lead to misjudgment and cannot guarantee timeliness.In recent years,artificial intelligence has developed rapidly,and researchers are gradually using deep learning techniques for crop pest and disease classification and detection.In this paper,we build on the above background and use convolutional neural networks to design a classification network and detection network that can rapidly identify crop diseases and insect pests,and deploy the classification network on the Android mobile terminal.The main work of this article is as follows:(1)To address the problems of large size and low recognition accuracy of convolutional neural network models for recognizing pest and disease images,an improved model DSC-LR-SE based on depth-separable convolution is proposed.Firstly,the DSC-LR structure is designed to pass the feature information from the convolutional layer into the feature fusion layer together with the original feature information after taking the negative operation,and introduces the residual connection and use the Leaky ReLU activation function.Then embeds the SE module to complete the final image classification.Among them,the DSC-LR structure solves the problem of ignoring negative feature value information during the training process of ordinary neural networks,and the SE module is used to enhance the useful feature channel information.The experiment shows that the DSC-LR-SE model has high accuracy in 59 types of public crop pest data sets,the number of network parameters is only 2.01 M,and the model weight is much smaller than the classical neural network model.(2)Based on the DSC-LR-SE model,the network structure is further optimized to design the hybrid attention module and the multiscale feature fusion module.The hybrid attention module enhances the weight of useful feature information and attenuates the weight of interference information such as noise by embedding channels and spatial attention mechanism,and the multi-scale feature fusion module extracts leaf disease features of different sizes by using its convolutional kernels of different scales.The improved model is named LMA-CNNs.experimental result shows that the accuracy of the LMA-CNNs model is 88.10% on the test set of 59 classes of public crop disease images,and the number of parameters is only 1.4M.By comparing the network models designed by different researchers using the same dataset premise,it is further verified that LMA-CNNs model not only possesses higher recognition accuracy but also has fewer parameters.(3)To address the problem of low detection efficiency and poor recognition accuracy of rice pests in the field,this article first collects related pictures from the Internet to build rice pest dataset;then improves the YOLOv4 algorithm and design the LMA-YOLOv4 detection model.This model uses LMA-CNNs as a trunk,uses the SPPF structure instead of the original SPP structure,and uses the Focal Loss loss function to equilibrate the positive and negative samples across the image.Finally,it is demonstrated that LMA-YOLOv4 can balance the accuracy of rice pest detection and recognition speed,and the model performance is excellent.(4)Based on DSC-LR-SE and LMA-CNNs crop disease recognition models,this paper designs a crop disease recognition system for Android mobile.The system supports crop disease recognition by uploading cell phone albums and shooting and uploading images,as well as real-time classification and prediction,and runs fast and easy to use. |
PDF Full Text Request