Incremental Object Detection Based On YOLOv5 And EWC Model

Current object detection technology is becoming increasingly sophisticated,with high detection accuracy and speed.However,people’s expectations for object detection technology go far beyond its accuracy and speed.People also hope that after obtaining a well-trained base model,object detection technology can continuously learn new unseen information,thereby achieving incremental object detection.In incremental object detection,the existing research techniques mainly rely on two-stage object detection algorithms.Although the two-stage object detection algorithm has an advantage in accuracy,it has a long training time and cannot meet the demand for real-time detection.Considering the shortcomings of the two-stage algorithm and finding that the accuracy and speed of the single-stage object detection algorithm can meet our needs,this paper uses the YOLOv5 model in the single-stage object detection as the base model for incremental object detection,and incorporates the incremental learning algorithm EWC to achieve incremental object detection.This article first introduces in detail the principles of the YOLOv5 model and EWC model.The YOLOv5 network structure consists of four parts: the input end,the main network,the Neck part,and the detection layer.The EWC algorithm is a posterior information-based algorithm that balances the importance of new and old parameters to achieve the purpose of remembering "old knowledge." To achieve incremental object detection,this paper combines the YOLOv5 model with the EWC algorithm,modifies the loss function of YOLOv5,adds the EWC algorithm as a regularization term in its loss function part,records the information of all learned data,and thus achieves incremental object detection.In the experiment,it was found that adding the EWC algorithm could not completely solve catastrophic forgetting.Later,it was found that it was caused by label conflicts between new and old datasets.To solve this problem,pseudo-label processing was added to the experiment,eventually achieving the purpose of incremental object detection.The experimental results show that the YOLOv5+EWC+pseudo-label model used in this paper performs well on the Pascal VOC2012 dataset.First,incremental learning is performed directly on the YOLOv5 model,and the results show that the YOLOv5 model can only remember the last learning result and recognize the data categories learned last time in the 5-15 one-time increment,10-5-5 two-time increment,and 5-5-5-5 three-time increment.Next,the YOLOv5+EWC algorithm is used for the three incremental experiments,and it is found that although the prediction accuracy of the YOLOv5+EWC algorithm on the old dataset is better than the YOLOv5 algorithm,its accuracy is still very poor.Then,pseudo-label processing is added in the data processing stage of the experiment.The experiment finds that the YOLOv5+EWC+pseudo-label model can remember the originally learned categories and continue to learn new data.The YOLOv5+EWC+pseudolabel model overcomes the catastrophic forgetting problem in incremental object detection,which has positive significance for the development of incremental object detection.The YOLOv5+EWC+pseudo-label model overcomes catastrophic forgetting in incremental learning,but the experimental results also reflect that its continuous learning times are limited.As the number of incremental learning increases,it becomes more difficult for the model to "remember" the information learned before,and learning new datasets also becomes more difficult.To break through the bottleneck of incremental object detection in terms of learning times,not only improvements in the algorithm are needed,but research on hardware devices is also indispensable.