A Method for Road Damage Detection Using Heterogeneous Data Based on Hybrid Kernel Attention

Front-view images and ground-penetrating radar (GPR) B-scans provide complementary observations for road damage assessment, covering both surface damage and near-surface structural anomalies. However, large discrepancies in the distribution, structural representation, and semantic characteristics of data hinder robust and scalable road damage detection across heterogeneous modalities, often leading to fragmented modality-specific pipelines. Meanwhile, dynamically varying multi-scale damage patterns, severe background clutter, and information loss during shallow-to-deep feature aggregation further complicate reliable detection. To address these issues, we propose HKANet, a shared detector architecture equipped with a hybrid kernel attention (HKA) mechanism that generates input-adaptive convolutional kernels by jointly modeling intra-kernel parameter interactions and inter-kernel importance. Based on HKA, HKANet incorporates three modules: multi-scale weighted attention convolution for shallow multi-scale feature extraction, a cross-stage partial network with dynamic spatial attention for clutter suppression and focused feature learning, and multi-scale global fusion for effective cross-level feature aggregation. Experiments on two front-view image datasets (USRDD and RDD2020) and a proprietary GPR dataset (RDRD) reveal that HKANet consistently outperforms the baseline, achieving up to 4.5% mAP@0.5 and 5.9% F1 improvements on USRDD/RDD2020 and 3.47% mAP@0.5 and 3.41% F1 gains on RDRD.