Pavement Crack Segmentation Using Dual-Encoder Feature Fusion Matching Network

Cracks constitute the most prevalent and recurrent damage pattern in pavement deterioration. They not only compromise surface integrity but also accelerate structural degradation, significantly shortening the operational lifespan of pavement infrastructure. The precise segmentation of cracks faces critical challenges due to morphological diversity, complex environmental backgrounds, and surface noise interference. To address these issues, this study proposes a dual-encoder feature fusion matching network integrating convolutional neural network (CNN) and Swin Transformer. The architecture employs parallel CNN-Transformer encoders to extract localized detail features and global contextual representations, respectively. A feature fusion module (FFM) enables multi-scale feature integration, while an innovative feature matching module (FMM) aligns decoder-encoder feature distributions to mitigate information loss during downsampling. The Transformer decoder progressively reconstructs high-resolution feature maps through skipconnection mechanisms. Evaluations on three benchmark datasets (DeepCrack, Crack500, and FIND) demonstrate state-of-the-art performance, achieving 83.07% Dice coefficient, 86.17 % F1 score, and 75.71 % Intersection over Union (IoU). The method exhibits robust segmentation capabilities under challenging scenarios including morphological variations, shadow occlusion, and noise contamination. This research advances intelligent detection technologies for pavement infrastructure maintenance.