Automatic Brittle Fracture Ratio Estimation Using Convolutional Neural Network Regression Based on Classmap Regulation

A convolutional neural network (CNN) based regression is proposed for estimating the brittle fracture ratio (BFR) in a image of drop weight tear test (DWTT) specimen. Different with previous complex semantic segmentation-based estimator, method extracts feature vector through global average pooling map and calculates BFR directly fully connected layer. By removing decoder network, number weights, training time, required GPU memory dramatically reduced. To train CNN, new loss function, which sum L1-norm between class activation ground truth inspection error, also designed. validate present method, images 1532, 79, 158 DWTT specimens obtained from real industrial site were used training, validation, test, respectively. The accuracy was evaluated on samples an error 5% or less divided by total samples, measure application. Despite having reduced weights inference time 85.8% 64.8%, respectively, has higher (96.2%) compared to that existing segmentation estimation (94.9%).