RepCrack: An efficient pavement crack segmentation method based on structural re-parameterization

The detection of pavement cracks is a significant challenge in the field of road maintenance. In response to this challenge, the industry has witnessed a growing trend towards the utilization of deep learning-based methods for crack detection. However, existing approaches utilize complex network structures which, while improving accuracy, compromise detection efficiency and hinder practical application. This paper proposes a novel pixel-level pavement crack segmentation model, RepCrack, which addresses the current limitations in accuracy and efficiency in pavement crack detection tasks. The proposed model is based on large kernel convolution and multi-scale feature extraction, in order to extract more detailed pavement crack information during the training phase. The model building process incorporates structural re-parameterization techniques, including the conversion of multi-branch structures to single on-path structures and the shortening of long sequence structures, which effectively reduces the complexity of the model. Furthermore, to enhance the model's capacity to extract features, an improved coordinate attention mechanism is incorporated to augment the network's focus on the crack region. To validate the detection effect, publicly available datasets, including the Crack Forest Dataset (CFD), German Asphalt Pavement Distress (GAPs) dataset, and Crack Tree dataset, were employed. The experimental results demonstrate that our proposed method exhibits a notable superiority over existing crack detection techniques. For instance, RepCrack, with a parameter count of 5. 8 million, attains an Intersection over Union (IoU) of 63. 8% on the CFD dataset and attains a state-of-the-art methodology (SOTA) performance, exhibiting a 5% enhancement in accuracy and a 68. 7% reduction in inference latency in comparison to SegNext.

Authors

• Minglun Ni
• Lei Chen
• Peixin Shi
• Ruiqi Ren

Keywords

• Pavement crack detection
• Semantic segmentation
• Structural re-parameterization
• Feature fusion
• Convolutional neural network
• Civil infrastructure