Abstract
Railway catenary system inspection is a critical task where high accuracy and reliability are essential to ensure operational efficiency and safety. This objective is achieved by assessing the technical condition of the infrastructure and maintaining a comprehensive inventory of its components. The application of machine learning methods to this problem is non-trivial, due to various constraints, including the cost of data acquisition. This paper presents innovative solutions leveraging domain knowledge to significantly improve the inference quality of machine learning models using existing training data. Key innovations include a two-stage approach and clustering of selected objects to extract regions of interest (ROI), dynamic confidence score weighting, and ROI masking, aimed at reducing false positives and enhancing precision. Additionally, the system was extended with ensemble learning methods and custom test-time augmentations (TTA). Proposed methods substantially improve metrics such as AP50, precision, recall, and F1-score, particularly in detecting small and hard-to-spot catenary components such as insulators. Notably, the proposed enhancements were optimized to mitigate processing time increases, enabling their application in industrial settings. The results demonstrate the effectiveness of integrating domain knowledge into the process of machine vision inspection, achieving an improvement in the F1-score metric from 61.97 (0.59) to 82.53 (0.38) compared to the baseline single-model approach while maintaining practical runtime constraints for real-world overhead catenary system inspection.