| Apple diseases directly affect the yield and quality of apples.Timely and effective disease identification can avoid improper control and pesticide abuse caused by misdiagnosis of diseases,and is the key to disease control.In this thesis,aiming at the identification of common apple leaf and fruit diseases,a disease identification model based on deep learning was constructed to realize the diagnosis of the diseased area of apple images.The specific research contents are as follows:（1）Apple leaf disease identification based on improved Faster R-CNNFor apple diseased leaves images,it is difficult to locate and identify these diseased leaves with small scale lesion and complex background in actual scenarios.To this end,we took five apple leaves diseases into account,and proposed an improved Faster R-CNN based apple diseased leaves detection method.Firstly,the training set data is expanded by data augmentation operation,and then the backbone is improved by using Res Nest and feature pyramid,and the Fast R-CNN was improved by further combining the cascade mechanism to optimize the generated proposal frame.The experimental results showed that the m AP@0.5 of the improved model reached 0.862,which could accurately identify small target lesions in complex backgrounds.（2）Apple leaf diseases identification based on lightweight YOLOv5sAiming at the massive occupation of the storage resources by the Faster RCNN,a lightweight YOLOv5 s based apple leaf disease identification method was proposed.The model used the Ghost structure and GD-PAN as the main body of the network to reduce the model complexity.Furthermore,it used the CIOU Loss as the boundary regression loss function and added the channel attention module to improve the model accuracy.The test results showed that the parameter size of the improved YOLOv5 s model was 3.69 MB,which is 539.31 MB lower than that of the Faster RCNN model,and its m AP@0.5 reached 0.872,which can accurately and effectively detect apple leaf diseases.（3）Apple fruit diseases identification based on GHTR2-YOLOv5 s and transfer learningIn view of the need to expand the disease field,a study on the data transferability of two similar disease fields of apple leaves and apple fruit was carried out.A disease identification model: GHTR2-YOLOv5 s was constructed with the goal of identifying four common apple fruit diseases,and transfer learning was performed using apple leaf disease data to provide support for the task of fruit disease identification.The construction process of GHTR2-YOLOv5 s is as follows: adding a phantom structure on the basis of YOLOv5 s and adjusting the width of the feature map to obtain a small baseline model,improving the accuracy of the model through the convolution block attention module and weighted bidirectional feature pyramid network,and using the TR2 module to enhance the model’s acquire ability of getting global information.After the improvement,the number of parameters of the model was 2.06 M,and the recognition speed reached 0.065s/sheet;its m AP@0.5 reached 0.909,and its m AP@0.5:0.95 reached 0.671.Through the combination of online image enhancement and secondary transfer learning,the study explored the feasibility of transferring from the field of leaf disease identification to the field of fruit disease identification.The m AP@0.5 of the model in this mode reached 0.916,which was 8.5% higher than the original sample model,proving the effectiveness of similar disease domain transfer. |
