Residual neural network-based fully convolutional network for microstructure segmentation

Junmyoung Jang

doi:10.1080/13621718.2019.1687635

Residual neural network-based fully convolutional network for microstructure segmentation

2019-11-07

25

PRE

AI

DOI：10.1080/13621718.2019.1687635

Original

Original request for help

Save

Abstract

En 中文

In this study, microstructures of weldment produced using carbon steel A516 grade 60 were analysed via a deep learning approach to measure the fraction of acicular ferrite which considerably influences on mechanical properties of carbon steel. The fully convolutional network was used to conduct the image segmentation. Submerged arc welding was used for welding, and the dataset was constructed using optical microscope. The model was compiled with ResNet, which is the state-of-the-art classifier used as an encoder. The model is trained to distinguish acicular ferrite from microstructures of dataset images and then estimate its accuracy. As a result, the mean intersection over union, which is a metric commonly used to evaluate image segmentation, was shown to be higher than 85%.

Keywords:

Submerged arc welding

carbon steel

acicular ferrite

fraction

segmentation

deep learning

fully convolutional network

ResNet

Residual neural network-based fully convolutional network for microstructure segmentation

Abstract

AI Summary

Journal

Organization