Pavement Crack Detection Based on the Improved Swin-Unet Model

Abstract

Accurate pavement surface crack detection is crucial for analyzing pavement survey data and the development of maintenance strategies. On the basis of Swin-Unet, this study develops the improved Swin-Unet (iSwin-Unet) model with the developed skip attention module and the residual Swin Transformer block. Based on the channel attention mechanism, the pavement crack region can be better captured while the crack feature channels can be assigned more weights. Taking advantage of the developed residual Swin Transformer block, the encoder architecture can globally model the pavement crack feature. Meanwhile, the crack feature information can be efficiently exchanged. To verify the pavement crack detection performance of the proposed model, we compare the training performance and visualization results with the other three models, which are Swin-Unet, Swin Transformer, and Unet, respectively. Three public benchmarks (CFD, Crack500, and CrackSC) have been adopted for the purpose of training, validation, and testing. Based on the test results, it can be found that the developed iSwin-Unet achieves a significant increase in mF1 score, mPrecision, and mRecall compared to the existing models, thereby establishing its efficacy in pavement crack detection and underlining its significant advancements over current methodologies.