From fragmented to flawless: A large-scale synthetic micrograph library for benchmarking microstructural image...

Area of Science:

• Materials Science
• Image Analysis
• Computational Materials Science

Background:

• Accurate grain boundary extraction is crucial for quantitative microstructural analysis.
• Experimental artifacts like noise and incomplete boundaries hinder manual analysis, making it time-consuming and subjective.

Purpose of the Study:

• To present a large-scale synthetic dataset for training and benchmarking deep learning models.
• To address automated noise cleaning and grain boundary reconstruction in metallographic images.

Main Methods:

• Generation of clean, single-phase polycrystalline micrographs using the polySim library.
• Programmatic degradation of images by adding noise and introducing grain boundary discontinuities.
• Creation of 14,999 paired PNG images for noise removal and grain boundary reconstruction tasks.

Main Results:

• A comprehensive dataset enabling the development of robust deep learning models.
• Facilitation of automated, objective, and reproducible microstructural image analysis.
• Overcoming the bottleneck of manual defect correction in metallographic image processing.

Conclusions:

• The synthetic dataset is a valuable resource for advancing automated microstructural analysis.
• Deep learning models trained on this dataset can significantly improve efficiency and objectivity.
• This work paves the way for more reliable quantitative analysis of material microstructures.