HKANet: An efficient software framework for cross-domain road damage detection using hybrid kernel attentionImage 999

Front-view images and ground-penetrating radar (GPR) B-scans provide complementary observations for the comprehensive assessment of road damage, capturing both surface distress and near-surface structural anomalies. However, deploying a detection system across these heterogeneous modalities is challenging due to domain shifts and multi-scale damage patterns. In this software impact paper, we introduce HKANet, an open-source, modular software framework for real-time, cross-domain road damage detection. A primary contribution is the implementation of a hybrid kernel attention (HKA) mechanism. Unlike static convolutions, this module dynamically aggregates multi-scale spatial features, adapting to both fine-grained surface cracks and subsurface anomalies. Built on PyTorch, the framework provides a cohesive programming interface for model training, evaluation, and deployment. This ensures reproducible evaluation on datasets including the road damage registration dataset (RDRD), the United States road damage dataset (USRDD), and road damage dataset 2020 (RDD2020). Benchmarks show that the core architecture maintains a computational footprint of approximately 16.5M parameters and 48.5 giga floating-point operations per second (GFLOPs), varying marginally based on the classification head. Available on Code Ocean, the software provides a scalable baseline for structural health monitoring applications.