HWCNet: A Semantic Segmentation Model for Highway Asphalt Pavement Cracks

To address challenges in the accuracy and integrity of semantic segmentation models for highway asphalt pavement crack detection, a novel model, highway crack net (HWCNet), is proposed, combining convolutional neural network (CNN) and transformer techniques. HWCNet features a custom backbone encoder integrating MobileOne blocks, dilated convolution, space-to-depth convolution (SPDConv), and an efficient multiscale attention (EMA) mechanism, with a transformer-based attention-based intrascale feature interaction (AIFI) module as the bottleneck. Its decoder utilizes dense upsampling convolution instead of traditional interpolated upsampling, alongside the adaptively spatial feature fusion (ASFF) structure. To address class imbalance, the focal Dice loss function is applied. Comprehensive comparisons across data sets with varying sample sizes, image qualities, and positive-negative sample ratios demonstrate the effectiveness of HWCNet. Results show that HWCNet outperforms UNet and other classical models, achieving F1-Scores of 88.65%, 76.73%, 88.78%, and 86.04%, representing improvements of 0.03%, 1.02%, 0.69%, and 3.98%, respectively. Unlike other algorithms with limited adaptability across diverse data sets, HWCNet performs well across data sets with different sample sizes, image qualities, and positive-negative sample contrasts. Notably, it excels in large-scale data sets with low contrast ratio of positive and negative samples and thin cracks, achieving an F1-scores improvement of up to 3.98%. HWCNet overcomes common limitations in segmentation accuracy, result completeness, and boundary differentiation, producing complete and reliable crack labels that provide a solid basis for maintenance decision-making. Furthermore, HWCNet adapts well to varying crack widths and complex backgrounds, exhibiting exceptional segmentation and generalization capabilities, making it well-suited for diverse pavement crack segmentation tasks.