A method for superfine pavement crack continuity detection based on topological loss

Abstract

Abstract Deep convolutional neural networks have become a popular tool for the automatic detection of pavement cracks. Despite their widespread use, the models currently available tend to emphasize pixel‐level classification accuracy for cracks, often overlooking the critical aspect of crack continuity. Addressing this gap, the authors’ research introduces a new method for the continuous detection of ultrafine pavement cracks, centred around the concept of topological loss. The authors’ novel approach hinges on expressing the disconnectivity between the background areas in an image through the connectivity of the cracks themselves. The proposed loss function accomplishes this by penalizing the unnecessary disconnection of the background areas, thereby minimizing the risk of false‐positive crack predictions. In this study, the authors trained and tested a crack‐detection network using the Crack500 dataset, as well as a new dataset specifically compiled for ultrafine crack analysis. The experimental outcomes indicate that the amalgamation of the mean squared error and the authors’ novel spatial topological loss function leads to a substantial improvement in the topological structure of ultrafine crack detection.

Keywords

• convolutional neural nets
• crack detection