Lightweight segmentation algorithm for real-time asphalt pavement crack detection on resource-constrained devices

Abstract

Abstract Transverse and longitudinal cracks are the typical and major pavement damage during routine asphalt pavement maintenance. Efficient and accurate detection of pavement cracks is crucial to the resource-constrained edge devices. Although current pavement crack segmentation models achieve high accuracy, inefficient feature extraction results in high computation and poor real-time performance on edge devices. Therefore, a lightweight instance segmentation model, named Asphalt Pavement Crack-YOLO11n-seg (APC-YOLO11n-seg), is proposed to balance detection accuracy and computational efficiency. The Re-parameterized Convolution (RepConv) module is introduced to enhance the lightweight feature extraction network HGNetv2, resulting in the improved version named RepLightConv-HGNetv2 (RLC-HGNetv2). As the backbone, this network significantly reduces model parameters and computational load, thereby enhancing inference speed. An Efficient High-level Screening Feature Pyramid Network (E-HSFPN) is designed to optimize the neck network. The Coordinate Attention (CoorA) and Dynamic Sampling (DySample) are utilized to reconstruct the feature selection and fusion modules in HS-FPN, enabling efficient and accurate multi-scale crack feature integration. A Lightweight Asymmetric Segmentation Head (LASH) is proposed to reduce computational complexity in the head while maintaining precise localization and segmentation capability. Experimental results on a custom dataset show that compared with the baseline model, APC-YOLO11n-seg reduces the Parameters, GFLOPs, and Model Size by 56.18%, 33.33%, and 50.52%, while increasing AP50 (Box) and AP50 (Mask) by 0.8% and 0.2%. Deployed on the Jetson Orin Nano Super 8GB development board, APC-YOLO11n-seg achieves an inference speed of 64.1 FPS, complying with the requirements for real-time crack detection in routine maintenance.