This paper presents the transfer learning model for studying the accuracy of railway surface defect detection and classification. The research focuses on one of the major structural faults defect, the railway track. The types of defects include cracking, flacking, shelling, and spalling, many of which were selected because of the distinctiveness of the defect. While many research studies classify defect, they are often not accurately classified using traditional machine learning methods due to the variability of defects and complexity of defect patterns. To solve the problem, this research demonstrates a pre-trained model, with a DenseNet121 architecture, to provide transfer learning to focus on surface fault detection. The model achieved accuracy of close to 99%, which is near perfect, and is significantly better than majority generalized models. The proposed model not only uses the representational strength of DenseNet but allows for good generalization across all defect classes with minimum feature engineering and augmentation. This framework can also be extended to other industrial fault detection tasks, which can demonstrate its scalability and effectiveness in the real-world

Roshni Mustafa Soho; Sanam Narejo; Muhammed Zakir Sheikh; Muhammad Khalid; Enrique Nava Baro; Agata Manolova

2025 28th International Symposium on Wireless Personal Multimedia Communications (WPMC)

https://ieeexplore.ieee.org/document/11351231

Railway surface defects; Artificial Intelligence; Transfer Learning; Image Processing