Green AI-Driven Synthetic Data Generation for Improved Industrial Metal Surface Defect Detection

Defect detection of industrial metal surfaces is often limited by the scarcity of annotated data and the high cost of manual labeling, which limits the performance of data-driven inspection systems. This paper presents a Green AI-driven framework for synthetic data generation that efficiently augments industrial defect datasets while minimizing computational and annotation costs. The proposed approach employs a deterministic procedural pipeline to generate synthetic defect images together with aligned segmentation masks at the pixel-level, providing accurate supervision without the computational overhead of deep generative models. The framework is evaluated on the Kolektor Surface-Defect Dataset 2 (KSDD2), which is extended into a hybrid dataset through synthetic augmentation followed by human-in-the-loop validation, to ensure the realism and reliability of the generated samples and annotations. A lightweight state-of-the-art anomaly detection model, SuperSimpleNet, is trained under controlled ablation settings on both the original and hybrid datasets using identical test splits for fair comparison. Performance is evaluated using both image-level and pixel-level metrics, including AUROC and average precision. Experimental results demonstrate that the proposed hybrid dataset consistently improves defect detection and localization performance while substantially reducing the need for costly data acquisition and manual annotation. Unlike computationally intensive generative approaches based on GANs or diffusion models, the proposed procedural framework synthesizes defects directly within predefined regions, enabling automatic generation of pixel-accurate ground-truth annotations without human intervention. The resulting hybrid dataset improves all evaluation metrics, achieving the largest gain in anomaly localization performance, with AP-loc increasing by 5.3%, while image-level AUROC increases from 0.9840 to 0.9907. By combining efficient synthetic data generation with precise automatic annotation, the proposed approach provides a resource-efficient solution that aligns with Green AI principles.