CDIF-Net: cross-dimensional interactive fusion network with dual-branch attention for pavement crack segmentation

Abstract Accurate detection of pavement cracks is critical for ensuring road safety and maintaining infrastructure integrity. While recent advances in deep learning have improved crack detection through effective local feature extraction, their limited receptive fields often hinder the modeling of long-range feature dependencies, which are essential for capturing the elongated and continuous structures of cracks across large image regions. To achieve precise segmentation, crack detection methods must effectively integrate both local details and global semantics. In this paper, we propose a cross-dimensional interactive fusion network (CDIF-Net), a novel architecture designed for pixel-level pavement crack semantic segmentation. The network comprises a main encoder branch to capture fine-scale details and an auxiliary encoder to extract global semantics. To enhance edge-level segmentation, we propose a Cross-Frequency Feature Integration module, which decomposes crack images into high- and low-frequency components, thereby sharpening the main encoder’s sensitivity to crack edges and contours. Additionally, to mitigate interference from complex backgrounds, a Foreground–Background Contrastive Decoupler module guides the auxiliary encoder to focus on crack-bearing foreground regions across multiple scales. Furthermore, a Rough Dual-Branch Attention module employs an upper- and lower-bound pooling strategy to efficiently fuse features from both branches, ensuring a balanced integration of global context and local detail. Experiments on three benchmark pavement crack datasets demonstrate that CDIF-Net outperforms state-of-the-art models.