RFDD: A High-Quality Dataset based on Metrology for Fastener Defect Detection in High-Speed Railways

This paper presents the Railway Fastener Defect Dataset (RFDD), a high-quality, metrology-oriented benchmark designed for intelligent inspection of high-speed railway infrastructure. Unlike existing datasets that mainly rely on cropped single-fastener images with limited defect categories, RFDD preserves full-scene railway inspection layouts and introduces engineering-oriented defect definitions based on measurable operational maintenance criteria. The dataset comprises 1,350 high-resolution full-scene images (2048 × 2021) containing more than 8,100 annotated fastener instances, including one Normal class and five representative defect categories: Missing, Inverted, Displaced, Deformed, and Fractured. The dataset is constructed using a systematic framework governed by six visual consistency principles spanning geometry, optics, boundary gradient, noise, texture, and occlusion, thereby ensuring metrological consistency and engineering realism. All images preserve the original multi-fastener spatial layouts and are provided with high-precision pixel-level semantic annotations. In addition, comprehensive technical validation was conducted using multiple state-of-the-art detection architectures. Experimental results demonstrate that RFDD provides a realistic, reliable, and challenging benchmark for evaluating advanced computer vision algorithms and intelligent sensing systems under practical railway inspection conditions.