Comparative evaluation of Detectron2 Mask R-CNN and YOLOV8-Seg for microstructural image analysis in high-temperature superalloys

Tools

Iwediba, Isaac Ifeanyi and Vorontsov, Vassili (2025) Comparative evaluation of Detectron2 Mask R-CNN and YOLOV8-Seg for microstructural image analysis in high-temperature superalloys. In: 2nd Conference on Artificial Intelligence in Materials Science and Engineering, 2025-11-18 - 2025-11-19, Ruhr University Bochum.

[thumbnail of Iwediba-Voronstov-AIMSE-2025-Comparative-evaluation-of-Detectron2-Mask-R-CNN-and-YOLOV8-Seg]
Preview
 Text. Filename: Iwediba-Voronstov-AIMSE-2025-Comparative-evaluation-of-Detectron2-Mask-R-CNN-and-YOLOV8-Seg.pdf
Accepted Author Manuscript
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 logo
Download (1MB)| Preview

Abstract

A dataset of microstructural images of a Ni-based superalloy labeled for the gamma-prime intermetallic phase was split into “train” and “test” sets. Both datasets were employed to evaluate two deep learning models for segmentation quality. The same preprocessing, image augmentations and data splits were used for both models, to compare them under fair conditions. The first model used was the Detectron2 Mask R-CNN. It is a two-staged framework which first generates potential regions and subsequently refines them as object masks[1, 2]. It achieved a 41.89% in terms of mean average precision (mAP@[0.50:0.95]) on the “test” data. There was a particularly high performance for images with large gamma-prime precipitates, with an average precision of 99.0%. This demonstrates the model’s capability for picking up large, high-contrast features. It had a lower performance when used on images with small features, with an average precision of 32.3%. These sources of weakness are prevalent in approaches using region proposals. The second method was YOLOv8-Seg. It is one-stage model that directly predicts and segments objects in a single shot. It was more memory efficient, faster to train and could be run in real time [3]. This makes it suitable for high-throughput pipelines, or use under the constraints of a slow hardware configuration. Analysis of the two datasets further proved that YOLOv8-Seg was well suited to segmentation of images with more varied precipitate distributions. The comparative analysis highlights an inherent trade-off. On the large gamma-prime regions, Detectron2 presented more accuracy. YOLOv8-Seg achieved faster training and inference, higher flexibility for precipitate-dense microstructures. These observations are also consistent with what is known from computer vision systems in general: two-stage models produce more precise, but less computationally efficient outputs, whereas one-stage models focus on speed instead [4, 5]. This work will assist in the determination of segmentation tools towards materials informatics in the context of metallurgical manufacture. It shows that precise segmentation is critical to the automation of phase, particle/precipitate size, and morphology quantification in high-temperature superalloys. Through analyzing the strengths and limitations of each approach, this work demonstrates case-specific rationale for choosing one segmentation model over another based on either accuracy or efficiency demand. This can be used to expedite metallurgical research and design of advanced manufacturing processes.

ORCID iDs

Iwediba, Isaac Ifeanyi and Vorontsov, Vassili ORCID logoORCID: https://orcid.org/0000-0002-1958-0602;
CORE (COnnecting REpositories)