Performance Analysis of Deep Learning Models for Microstructure Image Classification of Carbon Steel

The development of artificial intelligence over the last two decades has been rapid. These rapid advances in artificial intelligence have created several opportunities to develop and apply artificial intelligence to assist human work. This research developed a classification system based on microstructure images to classify types of carbon steel, which is divided into three categories: steel with low, medium, and high carbon content. This study, we attempt to compare the performance of deep learning using various transfer learning models to determine the most suitable model for classifying types of carbon steel based on their microstructure images. We compared the transfer learning models ResNet50, EfficientNet-B5, DenseNet-121, and VGGNet-19. The total dataset used for training the four transfer learning models comprised 17617 microstructure image data from three classes of carbon steel, with the data divided into 75% for training and 25% for validation. After training with the transfer learning models ResNet50, EfficientNet-B5, DenseNet-121, and VGGNet-19, the performance accuracy of each model reached 99.21% for ResNet50, 97.63% for EfficientNet-B5, 95.25 for DenseNet-121, and 96.62% for VGGNet-19. A uniform value of 10 epochs was used during model training. Based on the obtained data, we conclude that ResNet50 is the best model for classifying the types of carbon steel based on microstructure images. By using deep learning technology, it is envisaged that time and effort would be saved during the production process, increasing productivity.