Semi-supervised Rail Defect Detection from Imbalanced Image Data

Rail defect detection by video cameras has recently gained much attention in bothacademia and industry. Rail image data has two properties. It is highly imbalanced towards the non-defective class and it has a large number of unlabeled data samples available for semisupervised learning techniques. In this paper we investigate if positive defective candidates selected from the unlabeled data can help improve the balance between the two classes and gain performance on detecting a specic type of defects called Squats. We compare data sampling techniques as well and conclude that the semi-supervised techniques are a reasonable alternative for improving performance on applications such as rail track Squat detection from image data.