PAF-Net: A Progressive and Adaptive Fusion Network for Pavement Crack Segmentation

Automatic crack detection remains challenging due to factors such as irregular crack shapes and sizes, uneven illumination, complex backgrounds, and image noise. Deep learning has shown promise in computer vision for pixel-wise crack detection, but existing methods still suffer from limitations such as information loss, insufficient feature fusion, and semantic gap issues. To address these challenges, a novel pavement crack segmentation network, called PAF-Net, is proposed, which incorporates progressive and adaptive feature fusion. To mitigate information loss caused by feature downsampling, a progressive context fusion (PCF) block is introduced to capture context information from adjacent scales. To better capture strong features from local regions, a dual attention (DA) block is proposed that leverages both global and local context information, reducing the semantic gap issue. Furthermore, to achieve effective multi-scale feature fusion, a dynamic weight learning (DWL) block is proposed that enables efficient fusion of feature maps from different network layers. Additionally, a multi-scale input unit is incorporated to provide the proposed segmentation network with more contextual information. To evaluate the performance of PAF-Net, we conduct experiments using four common evaluation metrics and compare it with multiple mainstream segmentation models on three public datasets. The proposed PAF-Net demonstrates superior segmentation accuracy for pixel-level crack detection compared to other segmentation models, as evident from qualitative and quantitative experimental results.