Automatic Steel Microstructural Quantification by Convolutional Neural Networks

Abstract

The microstructural analysis of a material allows the complete characterization of its mechanical properties. Thus, the performance of a mechanical component depends heavily on the identification and quantification of its microstructural constituents. Currently, this process is still done mostly manually by experts, making it slow, very labor-intensive and inefficient. It is estimated that an experienced expert takes 15 minutes per image to perform the proper identification and quantification of microconstituents. Therefore, a computational tool could greatly assist to improve the performance in this task. However, since a microstructure can be a combination of different phases or constituents with complex substructures, their automatic quantification can be very hard and, as a result, there are few previous works dealing with this problem. Convolutional Neural Networks are promising for this type of application since recently this type of network has achieved great performance in complex applications of computational vision. In this work, we propose an automatic quantification of microstructural constituents of low carbon steel via Convolutional Neural Networks. Our dataset consists of 210 micrographs of low carbon steel, and this amount of images was increased through data augmentation techniques, resulting in a total of 672 samples for training. With regard to network architectures, we used the AlexNet trained from scratch and the VGG19 and Xception both pre-trained. The results showed that CNNs can quantify microstructures very effectively.The microstructural analysis of a material allows the complete characterization of its mechanical properties. Thus, the performance of a mechanical component depends heavily on the identification and quantification of its microstructural constituents. Currently, this process is still done mostly manually by experts, making it slow, very labor-intensive and inefficient. It is estimated that an experienced expert takes 15 minutes per image to perform the proper identification and quantification of microconstituents. Therefore, a computational tool could greatly assist to improve the performance in this task. However, since a microstructure can be a combination of different phases or constituents with complex substructures, their automatic quantification can be very hard and, as a result, there are few previous works dealing with this problem. Convolutional Neural Networks are promising for this type of application since recently this type of network has achieved great performance in complex applications of computational vision. In this work, we propose an automatic quantification of microstructural constituents of low carbon steel via Convolutional Neural Networks. Our dataset consists of 210 micrographs of low carbon steel, and this amount of images was increased through data augmentation techniques, resulting in a total of 672 samples for training. With regard to network architectures, we used the AlexNet trained from scratch and the VGG19 and Xception both pre-trained. The results showed that CNNs can quantify microstructures very effectively.