EUVIMEDEUVIMED entdecken
Crossref

Shape2Mask: A Weakly Supervised Instance Segmentation Method for Aluminum Alloy Precipitations Based on Prior Shape Knowledge

Mingchun Li, Naishi Feng, Qiang Wang, Dali Chen

IEEE Transactions on Instrumentation and Measurement · 2026 · 75 · 1-18

DOI: 10.1109/tim.2026.3676165 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Transactions on Instrumentation and Measurement, 2026. Verfasst von Mingchun Li und 3 weiteren Autor:innen. Die Arbeit nennt 69 Literaturangaben. Gefördert von: China Postdoctoral Science Foundation, Natural Science Foundation of Liaoning Province.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The size, distribution, and shape of precipitations in aluminum alloys critically influence their mechanical properties. Transmission electron microscopy (TEM) enables the analysis and measurement of these nanoscale precipitations. However, the inherent fuzziness and overlapping in TEM images necessitate annotation by domain experts. It not only increases the cost of pixel-level annotation but also poses significant challenges for segmentation algorithms. To this end, we present Shape2Mask, a novel instance segmentation framework that utilizes low-cost box supervisions to predict masks by parameterized shape modeling. First, it establishes a novel shape-based segmentation paradigm. Instead of pixel-wise prediction, it predicts shape parameters of precipitations, effectively handling fuzziness in TEM images. Second, we design a hierarchical module based on graph neural network (GNN) and recent Kolmogorov–Arnold networks (KANs). The former is employed to extract effective geometric shape representations, while the latter is used to fit nonlinear shape parameters. Furthermore, benefiting from the shape-based paradigm, we introduce a differentiable prior loss function, explicitly integrating domain knowledge into the deep model. Experiments demonstrate our method accurately segments three types of precipitations with an AP score of 29.5 under box supervision. Quantitative measurements show that Shape2Mask achieves errors as low as 1.2 nm and 9.8 nm2 in precipitation length and area, respectively. It significantly outperforms existing weakly supervised methods for both segmentation and measurement metrics. These results prove the method’s effectiveness not only in visual tasks but also in providing valuable quantitative measurements for potential materials science applications. Our source code and TEM image dataset are available at https://github.com/neulmc/Shape2Mask

Förderung

  • China Postdoctoral Science Foundation
  • Natural Science Foundation of Liaoning Province

Zitieren

APA

Li, M., Feng, N., Wang, Q., & Chen, D. (2026). Shape2Mask: A Weakly Supervised Instance Segmentation Method for Aluminum Alloy Precipitations Based on Prior Shape Knowledge. IEEE Transactions on Instrumentation and Measurement, 75, 1-18. https://doi.org/10.1109/tim.2026.3676165

Vancouver

Li M, Feng N, Wang Q, Chen D. Shape2Mask: A Weakly Supervised Instance Segmentation Method for Aluminum Alloy Precipitations Based on Prior Shape Knowledge. IEEE Transactions on Instrumentation and Measurement. 2026;75:1-18. doi:10.1109/tim.2026.3676165

BibTeX
@article{Li2026,
  author = {Li, Mingchun and Feng, Naishi and Wang, Qiang and Chen, Dali},
  title = {Shape2Mask: A Weakly Supervised Instance Segmentation Method for Aluminum Alloy Precipitations Based on Prior Shape Knowledge},
  journal = {IEEE Transactions on Instrumentation and Measurement},
  year = {2026},
  volume = {75},
  pages = {1--18},
  doi = {10.1109/tim.2026.3676165}
}
Kostenlos starten →