Semantic segmentation model for concrete cracks based on parallel Swin-CNNs framework

In recent years, crack detection has been the focus of relevant research since concrete fractures are the most dangerous damage to structures. Computer vision-based approaches are frequently employed for their distinct benefits. However, the crack segmentation model based only on convolutional neural networks (CNNs) is still inadequate in generalization because of its inherent bias produced by its low contextual understanding capacity. This research provides a framework PSC (Parallel Swin-CNNs) that employs a multi-scale feature fusion pyramid decoder to partition concrete cracks semantically using Swin Transformer and CNNs. This research evaluates classic CNN models U-Net, U-Net++, DeepLabV3, PSPNet, Feature Pyramid Network (FPN), and DeepCrack on two datasets. The proposed PSC model achieves the best crack segmentation results, with a maximum improvement of 36.57% in F1-score and 62.38% in Intersection over Union value on both datasets, a reduction in parameters of 2.95%–40.89% except for PSPNet. The proposed PSC model demonstrates versatile applicability across various scenarios, effectively overcoming interferences such as light shadows, oil stains, potholes, and textured surfaces while maintaining high computational efficiency.