Detection Technology of Several Pests and Diseases in Orchard Based on Federated Learning and Improved Faster R-CNN

Inhalt:
Orchard pest detection is facing the problems of insufficient data imbalance in different areas and low accuracy and efficiency in orchard pest detection under complex environment. This paper presents a new method for detecting plant diseases and insect pests in orchard based on Federated Learning and improved Faster R-CNN. Adding a restriction M to Federated Learning yields a detection model that integrates the data strengths of all parties. At the same time, ResNet-101 is used instead of VGG-16 of the original architecture to build the basic convolution layer. Then, the feature maps of different convolution layers are fused in multiple dimensions. Finally, a soft-NMS algorithm is presented after the RPN network. In order to meet the training needs of the model, enough data is obtained from the image collected by sample expansion method and relevant experiments are carried out. The experimental results show that the improved Faster RCNN can achieve 89.34% mAP in defective parts detection. The model training speed increased by 59% after using the Federated Learning mechanism. Compared with other detection algorithms under the same conditions, the comprehensive performance is the best, which can meet the needs of orchard pest detection and reduce the calculation cost.