Automated Indentation Geometry Detection for Hardness Measurement: A Vision-Based Nondestructive Evaluation Approach

Hardness testing is one of the most widely applied methods for evaluating the mechanical properties of metallic materials in manufacturing and quality control. Among standardized techniques, Brinell, Vickers, and Rockwell are most commonly used, all relying on precise measurement of indentation geometry. However, conventional evaluation methods still depend on manual or semi-automated readings, which are prone to operator error and unsuitable for industrial automation. This study introduces an automated framework for indentation geometry detection and measurement using two complementary strategies: a classical image processing pipeline based on OpenCV and a YOLOv۱۰n deep learning model. To replicate realistic industrial conditions, St۳۷ steel blocks with unpolished, roughly machined surfaces and random markings were tested under Brinell, Vickers, and Rockwell A hardness standards. A batch of microscope images was gathered, manually annotated, and subsequently expanded via data augmentation to obtain a diverse and reliable labeled image set. The classical pipeline employed brightness enhancement, thresholding, morphological filtering, and contour-based geometry extraction, while YOLOv۱۰n was trained on the augmented dataset and refined using geometry-based post-processing. The YOLOv۱۰n model achieved a mean average precision (detection accuracy) of ۹۴% and reduced mean absolute error by ۴۲–۵۱% compared with classical image processing (p < ۰.۰۵). Furthermore, it maintained ۹۶–۹۷% accuracy under brightness variation, rotation, and mirroring, whereas the classical approach deteriorated to ۷۹–۸۴%. YOLOv۱۰n provides a robust, precise, and scalable solution for automated hardness testing, supporting the advancement of automation in industrial quality control.

کلیدواژه ها:

Nondestructive evaluation ، Vision-Based NDE ، Material Characterization ، Automatic Measurement ، Hardness Indentation Detection ، Deep Learning