Comparative study on small sample tool wear classification based on YOLO26 and vision transformer

Abstract

In the process of machining, manual misjudgment is prone to occur in the detection of tool wear. To address this issue, this paper explores the performance of two models, YOLO26 and Vision Transformer (ViT), under the conditions of a small sample and binary classification tool wear dataset based on a deep learning classification model. The convergence curves, confusion matrices, and performance metrics of YOLO26n-cls, YOLO26m-cls, YOLO26x-cls, ViT-Base/16, ViT-Base/32, and ViT-Large/16 models are comprehensively compared. Among the YOLO26 series models, the YOLO26n-cls model has the best performance metrics, with an accuracy rate of 99.81%, a recall rate of 99.75%, an F1 score of 99.78%, 1.53M parameters, 3.2 GFLOPs, an inference delay of 0.5ms, 2000 FPS, 0.88G GPU usage, and a weight size of 3.2M. The parameter count and GPU usage of the YOLO26m-cls model are both higher than those of the YOLO26n-cls model, while the performance metrics of the YOLO26x-cls model are generally lower than those of the YOLO26n-cls model. Among the ViT series models, the performance metrics of the ViT-Base32 model and the ViT-Large16 model are comparable to those of the YOLO26n-cls model, but their parameter counts and GPU usage are much higher than those of the YOLO26 series models. The ViT-Base16 model has the lowest performance metrics. The experimental results show that under the conditions of small sample size and binary classification, the YOLO26 series models are overall superior to the ViT series models. Moreover, the YOLO26 models are more lightweight than the ViT models, and YOLO26n-cls is the optimal model for this working condition.