Semi-Conv-DETR: A railway ballast bed defect detection model integrating convolutional augmentation and semi-supervised DETR

Abstract

Railway ballast bed defects, including subsidence, mud pumping, and abnormal water, pose significant safety risks by destabilizing the railway ballast beds. Timely detection and repair of railway ballast bed defects are vital for safeguarding the security of both the trains and their passengers. Ground-Penetrating Radar (GPR) is widely used for railway ballast beds inspection and evaluation owing to its high speed and non-destructive characteristics. However, GPR image data contain considerable noise, and the distinct shapes and sizes of each ballast bed defect renders it challenging to apply a unified data annotation standard, which hampers the development of railway ballast bed defect detection models. Considering the distinct wave-like characteristics of GPR data and the vaguely contours of the defects to be identified, we propose a convolutional augmentation operation tailored for GPR images. Furthermore, we also investigate Semi-Supervised Learning by employing limited annotated railway ballast bed inspection data along with a vast amount of unlabeled data to joint train the DETR detection model. To sum up, we proposed a semi-supervised DETR model supplemented with convolutional augmentation for railway ballast bed defect detection, termed as Semi-Conv-DETR model. Experimental outcomes indicate that Semi-Conv-DETR shows an improvement of 58.6 % in accuracy when compared to the classical Faster-RCNN model.