Speeding up the annotation process in semantic segmentation industrial applications

This study addresses the significant bottleneck of manual data annotation in industrial materials science, specifically for semantic segmentation tasks. The authors compare two labeling strategies: traditional manual annotation from scratch and a semi-automatic approach where unsupervised algorithms generate initial masks that are subsequently refined by human experts. The researchers evaluated several unsupervised techniques—including multi-Otsu thresholding, superpixels, k-means clustering, a fully unsupervised convolutional neural network, and the Segment Anything Model (SAM)—to determine their efficiency in accelerating the labeling of high-resolution steel microstructure images.

The study demonstrates that integrating unsupervised pre-annotation into the workflow significantly reduces human labor. By using these algorithms to handle the bulk of the segmentation, the total annotation time for the dataset was reduced from 170 hours to 37 hours, representing a 78% increase in efficiency. The authors found that while no single unsupervised method was perfect, they provided high-quality starting points that allowed experts to focus on refining complex boundaries rather than drawing masks from scratch. This methodology enabled the creation of the largest publicly available, fully annotated steel microstructure dataset, consisting of 82 high-resolution images.

Data quality and availability are critical for the successful deployment of machine learning in industrial settings, yet the high cost of expert-level annotation often prevents the development of robust models. By providing a validated, efficient workflow for generating high-quality datasets, this research lowers the barrier for applying deep learning to materials characterization. The release of the MicroSteel dataset and the associated benchmark model provides a valuable resource for researchers to further develop and test segmentation models in a real-world industrial context.