A Preprocessing Method to Improve Edge Crack Detection from Railway Tunnel Lining Images

Abstract

Crack detection is critical for guaranteeing the safety of bridges, railways, and other infrastructures; however, it is a difficult task, particularly for tunnels. Tunnel lining images are primarily acquired using vision sensors, and cracks typically appear throughout an entire image. For crack detection using convolutional neural networks, the recognition accuracy is unsatisfactory when the cracks are at the edge of the image. Hence, an image preprocessing method is proposed to process railway tunnel data. In this method, the relative position of cracks in an image is changed by adding different sizes of borders to the crack images, and four different detection models are used for training to examine the effectiveness of the preprocessing method. Experimental results show that the proposed preprocessing method achieves better detection results for all four models. In the custom dataset, the border size is set to 1/9 of the original image size, which is the most effective size for improving edge crack recognition, where a maximum improvement of 8.4% compared with the control group is achieved. Additionally, black pixels (pixel value 0) are used to fill the border, which is better than using white pixels (pixel value 255).