Automatic Identification of the Multiphase Microstructures of Steels based on ASPP‐FCN

Abstract

The properties of steels are often closely related to the category and distribution of their microstructures. However, the classification and quantitative analysis of multiphase steel are mostly performed manually, which is time‐consuming and laborious. Moreover, due to the variant experience of experts, the classification results of one image will be different. In this article, an automatic classification model is proposed to identify the multiphase microstructures of steels by means of semantic segmentation, which realizes the automatic statistics of the relative content of each crystalline phase. Semantic segmentation can assign each pixel in an image to one of a set of predefined categories, which can accomplish two tasks concurrently: phase identification and segmentation. We propose an improved fully convolutional network ASPP‐FCN, and carry out comparative experiments with different networks, including FCN, DeepLab v3+, Unet, Enet and PSPnet. The final pixel accuracy of ASPP‐FCN in each category is more than 95%; and the mean intersection over union is over 80%, the results demonstrate that ASPP‐FCN has better segmentation effect in the automatic recognition of the multiphase microstructures of steels.

A Deep Learning‐Based Approach to the Estimation of Jominy Profile of Medium‐Carbon Quench Hardenable Steels

“…CNNs are a DL algorithm that is widely used in image [42] and video recognition, natural language processing, and other applications that require the processing of sequential data. They are intensively applied also in steel production for a quite wide range of tasks, such as, for instance, shape and surface defects detection and classification, [43][44][45] microstructure analysis and classification, [46,47] prediction of the end point of the converter, [48] and processing temperature data in continuous casting. [49] The key idea behind CNNs is to use a series of convolutional layers to extract features from an input image, followed by one or more fully connected layers to classify the image.…”