An automatic segmentation and quantification method for austenite and ferrite phases in duplex stainless steel based on deep learning

Traditional microstructural analysis and phase identification largely rely on manual efforts, inevitably affecting the consistency and accuracy of the results. Historically, the identification of ferrite and retained austenite phases and the extraction of grain information have predominantly been conducted by experts based on their experience. This process is not only time-consuming and labor-intensive but also prone to discrepancies due to the subjective nature of expert judgment. With the continuous advancement of deep learning technologies, new solutions for the classification and analysis of microstructures have emerged. This study proposes a microstructural segmentation method for dual-phase steel based on the Mask R-CNN deep learning model, which can quickly and accurately segment ferrite and retained austenite phases in dual-phase steel subjected to different heat treatment temperatures, enabling quantitative analysis of grain information. First, lightweight dual-phase steel is subjected to heat treatments at five different temperatures, and electron microscope images are obtained as training and testing data for the network. Through data preprocessing, annotation, and augmentation, a microstructural image dataset is constructed. Subsequently, the Mask R-CNN deep learning model is employed to recognize and segment the microstructural dataset of dual-phase steel. From the mask images output by the model, quantitative parameters such as the volume fraction and average grain size of ferrite and retained austenite are successfully extracted. Furthermore, the approach demonstrates high portability and applicability, particularly relying on a small sample dataset.