Dual-Domain Adaptive Downsampling and Hybrid Feature Integration for Lightweight Pavement Defect Detection With UAV Images

Pavement defect detection aims to accurately identify various surface defects and their unpredictable patterns while providing precise measurements, significantly enhancing road maintenance efficiency, safety, and traffic optimization. Utilizing an unmanned aerial vehicle (UAV) to capture road images can further reduce maintenance costs. However, existing detectors struggle with excessive feature loss and insufficient feature extraction when handling sparse pavement defect features, leading to low detection accuracy, while their high computational demands pose significant barriers to accurate detection in resource-constrained devices. To address these issues, we present a novel dual-domain adaptive downsampling and hybrid feature integration method (DDHI) for pavement defect detection, which is lightweight and highly accurate. First, we propose a dual-domain adaptive downsampling module (DADM) to reduce the loss of multiscale sparse defect features through downsampling in both the frequency and spatial domains while employing weighted fusion of different frequency-domain features and feature mapping in the spatial domain. Second, we propose a hybrid feature integration module (HFIM) to enrich the extracted sparse defect features by performing feature extraction sequentially in both the spatial and channel domains while considering both local and global feature extractions in the spatial domain. In addition, we propose a challenging UAV road defect dataset (UAVRD), which contains 1818 images, covering six defect categories and four scenes, having high density and multiple shadow conditions. Compared to the state-of-the-art real-time detection transformer (RT-DETR), our method reduces parameters and giga floating-point operations per second (GFLOPs) by 56% and 51%, respectively. Moreover, on the UAVRD and a publicly available dataset, our method achieves increases of 2.7% and 1.6% in mAP@.5 scores, respectively. The code and datasets are available at https://github.com/road-defect-detection/DDHI.