| Agriculture is the cornerstone of China’s economic construction and social development and has a direct impact on thousands of households.In the field of agriculture,crop diseases are the main factors that threaten crop biosecurity,affect food output and limit agricultural development.Therefore,scientific diagnosis of crop diseases and rapid response measures are the key to solve the disease problem and promote sustainable development of agriculture.Compared with the traditional disease recognition method,deep learning with high efficiency and fast identification of advantages has been widely applied.In this context,this paper proposes a crop disease identification method based on deep learning,first improving the residual network ResNet-50,then using the improved network model to study the identification of crop diseases,and finally achieving the target detection of the disease area through YOLOv3 to facilitate the subsequent control of crop diseases.The main work of this thesis is as follows:(1)The experimental data set is optimized to enrich and standardize the data.On the benchmark network ResNet-50,the effects of different gradient descent algorithms on crop disease image classification under different learning rates were tested,and the experimental results were compared and analyzed to select an optimizer and learning rate suitable for network training.(2)For the task of crop disease identification,this thesis constructed a hybrid attention module and combined it with ResNet-50 network to strengthen the network’s attention to important information and channels.And on this basis,the network is further improved,and a bilinear attention residual network based on feature fusion is designed by using bilinear idea and feature fusion of multi-level convolution layer.Compared with the benchmark network ResNet-50,the recognition accuracy of crop diseases by the improved network model is improved by 0.93%.(3)On the basis of improving the network,with the help of transfer learning,use the auxiliary data set to pre-train the network ResNet-50,save the trained pre-training model,and train the target data set through weight migration and network fine-tuning.And evaluate the network model through confusion matrix and feature map visualization,the experimental results show that the improved network has strong robustness and can effectively improve the classification ability of crop diseases,and its recognition accuracy is 1.78% higher than that of the benchmark network.(4)Detection of crop diseases using object detection algorithms.In this thesis,LabelImg software is used to manually annotate the disease information of crops,and the diseased area of crop disease is detected by YOLOv3 algorithm,and the experiment shows that in the detection of four diseases in apple leaf,the overall detection mean accuracy mAP reaches 83.78%.In summary,the crop disease identification method based on deep learning proposed in this thesis has high identification accuracy.Compared with traditional methods,it can obtain relatively ideal recognition results,which has certain practical significance in practical applications. |
