<p>Accurate segmentation of pavement cracks is crucial for ensuring road safety and longevity. However, existing segmentation networks often yield suboptimal detection results due to complex interfering factors in the background of pavement crack images, such as stains, road markings, and light variations, coupled with the high cost of acquiring large quantities of pixel-level annotated samples. To address these issues, this paper proposes a deep learning model based on dual-branch transitional feature coupling and hierarchical edge-guided fusion (DBTE).The model employs a multi-scale interactive adaptive coupling module, which dynamically fuses features of different scales through parallel branches, integrates local crack details with global structural information, and suppresses background noise. It introduces the Edge-Guided Feature Enhancement Module, which generates refined edge maps via EdgeMapEngine and strengthens edge features by combining with the attention mechanism of EdgeInformationInjector. Through a hierarchical fusion strategy, edge features and backbone network features are progressively integrated to enhance the ability of capturing details and discriminating structures. Experimental results based on the self-built urban pavement crack dataset from China University of Geosciences demonstrate that the proposed DBTE model achieves state-of-the-art segmentation performance on this core target scenario. Meanwhile, DBTE reaches top-3 competitive performance on two widely recognized public benchmarks (Crack500, CrackForest) without targeted optimization, verifying its excellent cross-scene generalization ability. The proposed model achieves an IoU of 85.79%, a Recall of 92.36%, and an F1-score of 91.83%, enabling more accurate pavement crack evaluation and analysis.</p>