Estimating Material Property Values from Fracture Surface Images with Vision Transformers

Abstract

Analyzing fracture surfaces to determine the type of fracture mechanics is important to improve the safety use of materials. Recent approaches apply deep neural networks to estimate fracture types or property values. This study aims to verify whether the combination of transfer learning with fine tuning and the vision transformer (ViT) model improves the accuracy of fracture surface image analysis. The verification results showed that the ViT with the attention mechanism displays superior performance to the convolutional neural networks (CNN) used in conventional fractography. It was also confirmed that the ViT particularly improves the systematic errors observed in conventional methods using CNN.