EfficientNet Architecture Family Analysis on Railway Track Defects

Abstract

Keeping railway tracks in correct conditions is of paramount importance to ensure adequate performance of trains, and to avoid any issues that might end up in accidents. The development of computational technology together with artificial vision techniques have boosted the use of video cameras for inspecting railway tracks. Some deep learning techniques have been applied to the day, being Convolutional Neural Networks the most popular ones. This paper presents the use of EfficientNet architecture on a railway track fault detection as a novel technique in this field. The paper compares B0 to B7 network families as well as studies the effect of input image resolution on a small dataset. On the one hand, results show that image resolution has an impact on validation accuracy and in fact, specific networks developed for particular image size are not always the best option for that image. B7 family has outperformed the rest of networks, reaching validation accuracy of 89.1%, while B2 scored 87.5% of accuracy, being the best in computational cost and performance ratio. In average, B7 and B2 have been proved to be the better solutions, with 84.6% and 84.2% accuracy respectively.