Accurate segmentation of high-density dislocations in micrographs with DeepCrack model

Abstract

Dislocations are fundamental defects in crystalline materials that critically influence mechanical properties and deformation mechanisms. Accurate segmentation of dislocation lines in micrographs is essential for quantitative microstructure analysis, yet it remains challenging due to complex contrast variations, high density, and morphological complexity. In this work, we adapt and retrain the DeepCrack model, a deep learning architecture originally designed for crack detection, to segment dislocations in experimental micrographs. Using a carefully annotated dataset of only 124 images acquired via scanning transmission electron microscopy and X-ray topography, which exhibit diverse dislocation morphologies and densities, the retrained model demonstrates strong generalization capability and accurate segmentation performance. It performs robustly even on highly challenging cases with extremely dense dislocation networks and significant image noise, albeit with a trade-off in output quality to avoid overfitting by shortening model training epochs. To overcome this limitation, a model recursion strategy was explored, which demonstrates potential for achieving top-tier recognition accuracy and high-quality binarized segmentation.