Attention-guided triplet framework for synthetic data generation and semantic segmentation of railway fasteners under data scarcity and disparity

Railway infrastructure monitoring faces significant challenges due to limited data availability, labor-intensive annotations, and imbalanced datasets—factors that hinder efficient fastener maintenance. To address these issues, this study proposes a novel three-phase framework that leverages attention-guided deep learning for the generation and analysis of synthetic railway fastener images. In the first phase, an attention-based Deep Convolutional Generative Adversarial Network (DCGAN) is introduced to generate high-fidelity synthetic images that closely mimic real-world conditions. Unlike conventional GANs, the attention mechanism enables the model to focus on critical structural features of fasteners, enhancing the realism and diversity of the generated data. The second phase applies advanced denoising techniques, with the DnCNN model outperforming traditional methods like Median Filtering in preserving fine details. The final phase employs a Convolutional Autoencoder (CAE) for accurate semantic segmentation, achieving 88.8 % accuracy on the synthetic dataset. This end-to-end methodology improves model generalizability, reduces reliance on manual labeling, and provides a cost-effective solution for automated railway inspection. By bridging the gap between real and synthetic data, it also lays the groundwork for scalable, intelligent infrastructure monitoring systems, supporting the advancement of safer and more efficient railway operations.