Learning Phase Representations for Microstructural Segmentation in Metallographic Images through Expert Knowledge
Keywords: Phase Fraction, Microstructure Segmentation, Material Segmentation
Abstract: Automated segmentation of metallographic images containing multiple phases such as martensite, ferrite, and pearlite is essential for quantifying different phases and thereby helping in the understanding properties of materials. Segmentation of these phases is challenging as they often exhibit overlapping boundaries, similar textures, and other more complexities that require a holistic understanding of the microstructures and correct phase representation within the image. To this end, we propose a novel approach for learning phase representations that captures the subtle differences between phases. Our proposed Phase Learning Module strategically integrates phase ratio information with image encodings to produce ratio-aware features that preserve critical spatial details. Materials scientists can roughly estimate phase ratios by examining an image, and our proposed model leverages this expertise. While we use expert-estimated phase ratios during inference, we train a model using accurate phase ratios obtained from target mask images. To our knowledge, this is the first use of class ratios as input in a deep learning segmentation model that serves as constraints to guide consistent phase proportions in predictions. Experimental results demonstrate segmentation performance improvements on both private and public datasets, with a 5.65% increase in Dice scores on the private dataset and a 6.48% improvement on the MetalDAM dataset with only 1.07% increase in model parameters. Furthermore, visualizations show that our approach leads to learning of more distinct and better phase representations across models. The code and private dataset will be made publicly available.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13852
Loading