FDCE-Net: A Frequency-domain Detail and Context Embedding Network for Pavement Crack Segmentation

Pavement crack segmentation is an important task in intelligent road inspection and maintenance. However, cracks often appear as elongated, discontinuous, and weak-boundary structures with a low foreground proportion. They are also easily affected by complex pavement textures, illumination changes, and noise, which makes fine-crack detection, continuity preservation, and background suppression challenging. To address these issues, this paper proposes FDCE-Net, a Frequency-domain Detail and Context Embedding Network for pavement crack segmentation. FDCE-Net adopts a convolutional neural network as the local feature extraction backbone and introduces a Frequency-domain Aware Detail Block (FADB), which enhances crack edges and high-frequency details from both frequency-domain and spatial-domain perspectives, thereby improving the perception of fine and weak cracks. Meanwhile, a global context modeling branch based on a state-space model is introduced to capture long-range dependencies among crack regions, and a hierarchical progressive structure is used to strengthen the representation of multi-scale crack morphology. Furthermore, a Global Feature Embedding Block (GFEB) is designed to adaptively embed global context into local detail features, enabling the integration of local textures, boundary details, and overall structural information. Experimental results on multiple public crack datasets show that FDCE-Net achieves better segmentation performance than representative comparison methods. On the DeepCrack dataset, FDCE-Net obtains an F1-score of 86.48\% and an IoU of 76.19\%, outperforming the second-best model CDS-Net by 0.52 and 0.81 percentage points, respectively. These results indicate that frequency-domain detail enhancement and global context embedding can improve the accuracy and completeness of crack segmentation in complex pavement scenes.