Detection of Rape Root Swelling Based on Improved YOLOV3 Model

Abstract

In view of the unsatisfactory recognition ability of target detection network for brassica napus root swelling in farmland, this paper proposed an improved YOLOV3 network model for automatic detection of brassica napus root swelling in farmland. Firstly, the data set of root swelling was improved, and the diversity of the data set was enhanced by data image enhancement. Secondly, the loss function of the original YOLOv3 network prediction model is improved to improve the accuracy of network model prediction. The generalization ability of the network model was enhanced by optimizing the activation function of the model. Finally, the improved YOLOv3 network model was used to conduct iterative training on the rape root swelling data set, and a better classification prediction network model of rape root swelling disease was obtained. Combined with the experimental results, the accuracy of the improved YOLOv3 network model was as high as 93.4%, and the average detection accuracy was 1.1% higher than that of the original YOLOv3 network model, which provided a better method for real-time detection of rape root tumor disease grade.