A Hybrid Transformer Framework with Receptive Field Context Aggregation and Learnable Basis Functions for Road Crack Detection

Abstract Road crack detection is crucial for infrastructure maintenance, enabling early identification of structural defects, thereby helping to prevent accidents. Despite the progress of deep learning in automating detection, current methods still fail to accurately characterize fine-grained cracks owing to limitations in extracting features across diverse receptive fields. Thus, an innovative hybrid detector named RC-DETR is proposed, aiming to address the aforementioned limitations. RC-DETR adopts the Real-Time Detection Transformer (RT-DETR) as the baseline and designs a Receptive Field Context Aggregation Architecture (RFCAA) that simultaneously captures CNN-processed local textures, aggregates hierarchical features across multiple resolutions, and obtains global contextual information, thus significantly improving detection precision for micro-cracks and textured defects. Furthermore, a Hierarchical Feature Refinement Encoder (HFRE) is proposed to suppress non-critical semantic information in multi-scale feature maps via parallel multi-branch processing, effectively enhancing discriminative crack signatures and improving localization robustness in complex backgrounds. Additionally, compared to the Multi-Layer Perceptron (MLP) which uses fixed activation functions, the RC-DETR framework employs a Kolmogorov-Arnold network (KAN), where the weights on edges are parameterized by learnable B-spline functions. This modification enables the RC-DETR framework to achieve higher accuracy with fewer parameters, significantly outperforming MLP-based approaches in fine-grained feature fusion and classification tasks. Experiments on the UAV-PDD2023 dataset show that RC-DETR achieves 85.5% mAP0.5 (an 11.5% improvement over RT-DETR) and 49.3% mAP0.5:0.95, with real-time inference at 10.3 ms on edge devices. Meanwhile, it reduces parameters by 6.7% and computational complexity (Giga Floating Point Operations Per Second, GFLOPs) by 26.5%. RC-DETR outperforms mainstream algorithms in balancing accuracy and efficiency for road crack detection, demonstrating superior robustness in complex environments. This offers a practical solution for automated road maintenance systems, particularly under diverse concrete pavement conditions.