RailNet: Railway Track Anomaly Detection via Image Processing With Hybrid Deep Learning Techniques

ABSTRACT Reliable detection of railway track faults is essential for preventive maintenance and safety. We introduce RailNet, a deep learning–based convolutional architecture that uses a DenseNet121 backbone as its feature extractor together with a streamlined classification head tailored to rail imagery. RailNet is fine‐tuned with a task‐aware augmentation policy designed to mimic in‐field conditions (viewpoint change, illumination variation, and occlusions) and includes built‐in gradient‐weighted class activation mapping (Grad‐CAM) interpretability to highlight defect regions that drive predictions. An ablation study quantifies the contribution of key head components (batch normalisation, dropout, and layer depth) to generalisation. RailNet is evaluated on a labelled dataset of faulty and non‐faulty track images; it achieves 96% accuracy with macro‐averaged precision, recall, and F1‐score of 0.96 on a held‐out test set, indicating balanced performance across classes. Heatmap visualisations consistently localise cracks and misalignments, supporting operator trust and triage. By combining a strong backbone with domain‐specific augmentation, quantified architectural choices, and built‐in interpretability, RailNet provides a reliable and efficient basis for early, automated track‐fault detection, enabling more proactive maintenance scheduling and contributing to reduced accident risk.