Pixel‐level thin crack detection on road surface using convolutional neural network for severely imbalanced data

Abstract Because roads are the major backbone of the transportation network, research about crack detection on road surfaces has been popular in computer and infrastructure engineering. When training a convolutional neural network (CNN) for pixel‐level road crack detection, three common challenges include (1) the data are severely imbalanced, (2) crack pixels can be easily confused with normal road texture and other visual noises, and (3) there are many unexplainable characteristics regarding the CNN itself. When it comes to very fine and thin cracks, these challenges are exaggerated and a new challenge is introduced, as there can be a discrepancy between the actual width and the annotated width of a crack. To tackle all these challenges of thin crack detection, this paper proposes a new variant of CNN named ThinCrack U‐Net, designed to provide thin results upon pixel‐level crack detection on road surfaces. The main contribution is to demystify how pixel‐level thin crack detection results are affected by different loss functions as well as various combinations of the U‐Net components. The experimental results show that ThinCrack U‐Net yields a significant performance boost in CrackTree260, from 65.71% to 94.48% F‐measure, compared to the existing variant of U‐Net previously proposed in the context of pixel‐level thin crack detection. Finally, this paper locates the source of undesirable result thickness and solves it with the balanced usage of downsampling/upsampling layers and atrous convolution. Unlike suggested by previous works, different loss functions show no significant impact on ThinCrack U‐Net, whereas normalization layers are proved crucial in pixel‐level thin crack detection.