Abstract
An efficient deep learning method is presented for distinguishing microstructures of a low carbon steel. There have been numerous endeavors to reproduce the human capability of perceptually classifying different textures using machine learning methods, but this is still very challenging owing to the need for a vast labeled image dataset. In this study, we introduce an unsupervised machine learning technique based on convolutional neural networks and a superpixel algorithm for the segmentation of a low-carbon steel microstructure without the need for labeled images. The effectiveness of the method is demonstrated with optical microscopy images of steel microstructures having different patterns taken at different resolutions. In addition, several evaluation criteria for unsupervised segmentation results are investigated along with the hyperparameter optimization.
Highlights
An efficient deep learning method is presented for distinguishing microstructures of a low carbon steel
The phases of steel alloys have been classified by the manual analysis of light optical microscopy (LOM) or scanning electron microscopy (SEM) images[4]
We demonstrated the segmentation of the microstructure of a low-carbon steel without labeled images using a deep learning method
Summary
An efficient deep learning method is presented for distinguishing microstructures of a low carbon steel. We introduce an unsupervised machine learning technique based on convolutional neural networks and a superpixel algorithm for the segmentation of a low-carbon steel microstructure without the need for labeled images. Further applications of SVM were presented by Gola et al.[7,8], where the microstructures of steel alloys were classified into constituent phases Another frequently applied technique is the random forest, which is a classification algorithm composed of multiple decision trees[9]. Deep learning is a class of machine learning algorithms that use multiple processing layers to learn representations of the raw input[13] These multilayers are called neural networks and are being actively applied to microstructure classification tasks. The width and height of the input image should be fixed to 224 × 224 pixels when implementing well-known networks such as V GGnet17, DenseNet[19], and ResNet1821
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
