Automatic Annotation of Crop Spot Image Based on Improved YOLOv5 Model
Image based object detection model research often required a large number of labeled image datasets for training and testing,but the existing manual annotation methods had problems such as low annotation quality,time-consuming and laborious,which seriously affected the modeling effect.In order to solve this problem,an improved image automatic annotation model YOLOv5-TR-BiFPN was proposed based on YOLOv5 model,aiming at automatic annotation of leaf and stem disease spots in rice images.Firstly,BiFPN(Bidirectional Feature Pyramid Network)was introduced into YOLOv5 structure to enable the model to integrate more features when the computational load was similar.Secondly,the ViT(Vision Transformer)module was used to enhance the model's target positioning capability,and the calculation methods of loss function and non-maximum suppression were further optimized to obtain a more accurate target frame.The research results showed that the average precision of YOLOv5-TR-BiFPN model for plant disease spot images reached 73%,which was 3%higher than YOLOv5s model.Using a small number of rice stem and leaf disease spot images to verify,the average precision of model training average mAP(mean Average Precision)reached 89.3%,which showed that YOLOv5-TR-BiFPN model could accurately label rice stem and leaf disease spots,achieve automatic labeling of crop disease spot images,and the labeling effect was good.
万方数据
维普
