Optimized Transfer Learning Models for Rail Surface Defect Classification with Explainable AI Validation
Abstract
Rail surface defects such as flaking, spalling, and squat are critical indicators of railway track degradation and require reliable and efficient automated inspection systems to ensure operational safety. This study presents a comparative evaluation of seven state-of-the-art pre-trained Convolutional Neural Network (CNN) architectures for rail surface defect classification, namely Inception-V3, MobileNet-V1, MobileNet-V2, MobileNet-V3, NasNetMobile, ResNet50, and EfficientNet-B0. Transfer learning was employed by freezing the convolutional backbones and optimizing the classifier head using Optuna-based hyperparameter search. The models were trained and evaluated on a benchmark rail surface defect dataset containing balanced samples of the three defect classes after redundancy reduction and class equalization. Performance was assessed using quality metrics, whereas computational efficiency was analyzed using Floating-Point Operations (FLOPs) and parameter complexity. Experimental results show that MobileNet-V2 achieves the highest classification accuracy of 81.6% with 599M FLOPs, whereas MobileNet-V3 achieves competitive accuracy of 78.5% with the lowest computational cost of only 132M FLOPs. Inception-V3 also demonstrates strong performance with 80.8% accuracy but requires substantially higher computational complexity (11.4B FLOPs). Pareto analysis confirmed that both MobileNet variants provide the best efficiency–accuracy trade-off among all evaluated models. To improve interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to the Pareto-optimal models. The visualizations revealed that MobileNet-V2 generally focuses more consistently on defect-relevant regions, particularly for flaking defects. The findings highlight the importance of combining performance benchmarking with Explainable Artificial Intelligence (XAI) analysis for safety-critical railway monitoring applications.