Adaptive Railway Track Inspection: A Semi-Supervised Self-Training Framework for Component and Defect Analysis

Railway infrastructure integrity is key for safe and efficient transportation; however, traditional track inspections rely on labour-intensive and error-prone methods. This work addresses these limitations by presenting an automated railway track inspection framework that combines a Convolutional Neural Network (CNN) based object detector with spatial reasoning. The system leverages multi-view data from an Automated Track Measuring Unit (ATMU) to accurately identify the five essential track components and infers three critical structural defects using spatial rule-based reasoning. Furthermore, the semi-supervised self-training mechanism utilizes spatially validated pseudo-labelling from high-confidence predictions, enabling adaptation to new collections and environmental conditions. Rather than proposing a new detector architecture, the contribution is a domain-specific formulation that couples component detection, spatial defect inference without defect-class supervision, and spatially validated self-training into one adaptive inspection pipeline. Comprehensive experiments on Mississauga collections, including snow, demonstrate improved detection accuracy and support the framework's practical value under the evaluated operating conditions.