Dislocations are line defects within crystal structures and have a critical influence on the mechanical properties of materials. Their presence, density, orientation, and distribution directly affect how materials deform, strengthen, and fail under mechanical loads. Electron Channeling Contrast Imaging (ECCI) enables visualization of dislocations over large areas in bulk samples, offering a promising alternative to Transmission Electron Microscopy (TEM). However, quantitative analysis of dislocations in ECCI images is still often performed manually, limiting throughput and reproducibility in microstructural characterization.

This work introduces an automated pipeline that combines deep learning with crystallographic analysis to detect, classify, and quantify dislocations in ECCI images. We use a YOLOv11-based segmentation model, initially pre-trained on TEM dislocation images (with contrast inverted to match ECCI appearance), and fine-tuned on a curated dataset of manually annotated ECCI micrographs. A tile-based sampling strategy with overlapping regions ensures robustness to varying image quality and local contrast.

Following segmentation, several post-processing steps are applied: skeletonization reduces dislocation masks to central lines; endpoint bridging connects nearby fragments to improve continuity; and overlapping segments from tiled regions are resolved by retaining the longest or most complete ones. Each resulting skeleton is assigned a unique identifier using connected component labeling, and small artifacts are filtered out based on pixel length.

The workflow is demonstrated in a model material system, a single-phase ferrite with a body-centered cubic (BCC) crystal structure. For this study, we assume that most dislocations are screw character, such that their line vector is parallel to their Burgers vector. Based on this assumption, and using crystallographic orientation data (Euler angles), standard BCC slip directions ($[111]$, $[\bar{1}11]$, $[1\bar{1}1]$, and $[11\bar{1}]$) are projected into the image plane. Each dislocation is classified by comparing its 2D orientation to these projected directions, within a tunable angular tolerance. Dislocations that deviate significantly or show curvature are labeled as “unclassified” or “curved.”

The pipeline outputs include overlaid visualizations, directional maps, and statistics containing both summary metrics (count, mean, total length) and detailed per-dislocation data. This enables high-throughput, reproducible analysis of dislocation structures without manual tracing.

By integrating crystallographic knowledge with modern computer vision, this work addresses a key challenge in microstructural analysis workflows and lays the groundwork for scalable, quantitative dislocation characterization in alloy development and deformation studies.