AI Driven Image Evaluation in Hardness Testing: Integration and Validation of the QAI Module within Qpix2

Abstract Hardness testing is an established method for evaluating mechanical properties of metallic materials in industrial production and research. While test execution has become highly automated, the evaluation of indentation images still introduces variability due to operator dependency and limitations of rule-based image processing methods. This manuscript presents the integration and validation of an artificial intelligence (AI) module for automated image evaluation (QAI) within the Qpix2 hardness testing system. The module is based on a convolutional neural network trained on more than 55,000 annotated impressions across Vickers, Knoop, and Brinell methods. The study describes the dataset composition, training and validation procedures, and baseline comparisons against manual operator evaluation and classical computer vision approaches. Experimental results demonstrate that AI-based detection achieves high recall and precision across varying surface conditions, including etched and low-contrast samples. Repeatability analysis shows reduced measurement variability compared to manual evaluation, and throughput studies confirm substantial time savings in batch testing. A dedicated comparative study with 100 Vickers HV1 impressions was additionally designed to investigate accuracy, repeatability, and evaluation time across manual, classical, and AI-based methods. The findings indicate that AI-supported evaluation improves robustness and traceability of hardness testing workflows while reducing operator dependency. The approach aligns with requirements for reproducibility, standardization, and digital documentation in safety-critical industries.