Structural continuity-aware pavement crack segmentation in unmanned aerial vehicle images with state-space feature interaction and dynamic global large-kernel attention

Deep learning is an important artificial intelligence technique for rapid and accurate crack segmentation from unmanned aerial vehicle (UAV) pavement images, supporting automated road inspection and maintenance. However, UAV images often contain scale variation, viewpoint changes, shadows, motion blur, and complex pavement textures, making thin cracks difficult to segment continuously and precisely. To address these challenges, this paper proposes an improved U-shaped encoder-decoder semantic segmentation network for UAV pavement crack extraction. In the encoder, a state-space-based feature interaction module and coordinate attention are introduced to enhance long-range crack structure modeling and directional-spatial representation. In the bottleneck and decoder, atrous spatial pyramid pooling, dynamic global large-kernel attention, and multi-scale channel attention are used to strengthen multi-scale context aggregation, local detail recovery, and background interference suppression. A UAV pavement crack dataset containing 7934 images is constructed from two image sources. Ablation experiments show that the proposed components provide complementary improvements. Compared with the baseline U-shaped encoder-decoder network, the proposed network improves the Dice coefficient, intersection over union, mean intersection over union, and Kappa coefficient by 2.18, 2.84, 1.45, and 2.38 percentage points, respectively. Same-dataset and external cross-dataset experiments further show that the proposed method maintains strong segmentation performance across the evaluated image sources, including an independent UAV-acquired pavement dataset, particularly in crack continuity preservation and boundary recovery.