Railway Track Defect Detection using Transfer Learning With EfficientNetB3

Regular railway track inspection is critical for ensuring safe and dependable train operations. Deformations, sediment issues, rail discontinuity, loose nuts and bolts, burnt wheels, superelevation, and misalignment developed on the rails as an account of inadequate maintenance, hasty investigations, and delayed detection pose a serious threat to the safe operation of rail transport. The traditional method of manually inspecting the rail track with a railway cart is inefficient and susceptible to human error and bias. In a country like India, where train accidents have claimed many lives, automating such approaches to avoid such accidents and save countless lives is not relatively rare. This study aims to contribute to the detection of cracks and interruptions in railways. This paper proposes an EfficientNet-B3 Model solution based on the pretrained EfficientNet-B3 DNN. A specialized branch has been incorporated into the network layer to compute the required weights, and the model possesses around 12 million parameters. These weights are automatically learned by training the entire DNN model end-to-end with the backpropagation algorithm. So the network learns to effectively detect the railway track zones while suppressing regions irrelevant to the classification. The detection mechanism achieves an accuracy of 93.55% while classifying a dataset of images.