MatSSL: Robust Self-Supervised Representation Learning for Metallographic Image Segmentation

TL;DR

MatSSL introduces a self-supervised learning approach with Gated Feature Fusion for metallographic image segmentation, enabling effective domain adaptation with limited unlabeled data and outperforming existing methods on key benchmarks.

Contribution

The paper presents a novel SSL architecture that effectively leverages small unlabeled datasets for metallographic segmentation, outperforming supervised and other SSL methods.

Findings

Achieves 69.13% mIoU on MetalDAM, surpassing ImageNet-pretrained models.

Provides up to 40% improvement in mIoU on EBC dataset over MicroNet-based models.

Enables effective domain adaptation with limited unlabeled data.

Abstract

MatSSL is a streamlined self-supervised learning (SSL) architecture that employs Gated Feature Fusion at each stage of the backbone to integrate multi-level representations effectively. Current micrograph analysis of metallic materials relies on supervised methods, which require retraining for each new dataset and often perform inconsistently with only a few labeled samples. While SSL offers a promising alternative by leveraging unlabeled data, most existing methods still depend on large-scale datasets to be effective. MatSSL is designed to overcome this limitation. We first perform self-supervised pretraining on a small-scale, unlabeled dataset and then fine-tune the model on multiple benchmark datasets. The resulting segmentation models achieve 69.13% mIoU on MetalDAM, outperforming the 66.73% achieved by an ImageNet-pretrained encoder, and delivers consistently up to nearly 40% improvement in average mIoU on the Environmental Barrier Coating benchmark dataset (EBC) compared to models pretrained with MicroNet. This suggests that MatSSL enables effective adaptation to the metallographic domain using only a small amount of unlabeled data, while preserving the rich and transferable features learned from large-scale pretraining on natural images.