Szczegóły publikacji
Opis bibliograficzny
AI-driven analysis of fracture types in metallic materials: classification, segmentation, and quantification with deep neural networks / Anna WÓJCICKA, Mateusz WOJTULEWICZ, Andrzej BRODZICKI, Magdalena Leonkiewicz, Patryk Maciejewski, Joanna JAWOREK-KORJAKOWSKA, Krzysztof Mroczka // W: ICCV-W 2025 [Dokument elektroniczny] : 2025 IEEE/CVF International Conference on Computer Vision Workshops : 19–20 October 2025, Honolulu, United States : proceedings. — Wersja do Windows. — Dane tekstowe. — Piscataway : IEEE, cop. 2025. — ( IEEE International Conference on Computer Vision Workshops ; ISSN 2473-9936 ). — Print on Demand (PoD) ISBN: 979-8-3315-8989-9. — e-ISBN: 979-8-3315-8988-2. — S. 3700–3708. — Wymagania systemowe: Adobe Reader. — Bibliogr. s. 3707–3708, Abstr. — Publikacja dostępna online od: 2026-02-23
Autorzy (7)
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 167849 |
|---|---|
| Data dodania do BaDAP | 2026-06-16 |
| Tekst źródłowy | URL |
| DOI | 10.1109/ICCVW69036.2025.00386 |
| Rok publikacji | 2025 |
| Typ publikacji | materiały konferencyjne (aut.) |
| Otwarty dostęp |  |
| Wydawca | Institute of Electrical and Electronics Engineers (IEEE) |
| Konferencja | IEEE International Conference on Computer Vision 2025 |
| Czasopismo/seria | IEEE International Conference on Computer Vision Workshops |
Abstract
Identifying and classifying fracture types in metallic materials is crucial for understanding failure mechanisms, predicting structural integrity, and improving material design. Accurate fracture analysis supports quality control, lifetime assessment, and the development of more resilient metal components in critical applications such as aerospace, automotive, and infrastructure. We propose a deep learning-based framework for automatic classification, segmentation, and quantitative assessment of fracture types in metallic materials. The method distinguishes brittle, ductile, and mixed fractures using microscopic SEM images and employs state-of-the-art convolutional neural networks, including EfficientNet and ConvNeXt, within a transfer learning pipeline. To address the class imbalance and the limited size of the dataset, we implement extensive data augmentation and regularization techniques. The best performing model (EfficientNet-B5) achieves an accuracy of 90.7% and an F1-score of 0.858. In addition to classification, we introduce a novel segmentation-based approach to estimate the level of mixedness on fracture surfaces, utilizing U-Net-like architectures and supported by optimized thresholding and test-time enhancement. This quantitative mixedness assessment enables deeper insight into fracture mechanics and energy absorption processes. The proposed solution demonstrates state-of-the-art performance in both classification and fine-grained fracture composition analysis, offering significant potential for applications in material diagnostics and structural failure analysis.
