AnisoNet: A Detecting Model for Initial Anisotropic Cracks in Complex Road Scenarios

Abstract

As a critical component of Intelligent Transportation Systems (ITS), automated pavement health monitoring plays a vital role in ensuring traffic safety and enabling predictive maintenance. At present, the accurate and efficient detection of incipient pavement cracks remains a key technical challenge in this field. However, challenges such as low resolution of captured road images and the small proportion of initial cracks result in current mainstream crack detection models exhibiting low recognition accuracy and a high likelihood of missed detections. To address these issues, this paper proposes a detection model called AnisoNet, specifically designed for early-stage cracks with anisotropic characteristics, aiming to enhance both the efficiency and accuracy of early crack detection. AnisoNet incorporates several innovative components: a three-level routing attention (TRA) mechanism, which dynamically adjusts the attention granularity to enhance the sensitivity of the model to small features; a multi-head routing dynamic convolution module (MRDC) to improve the ability to adapt to complex global features and curvilinear local features; a snake-aware receptive field convolution (SARFConv) designed to strengthen the model’s ability to detect subtle spatial variations and misalignments; and a multi-scale large kernel attention mechanism (LKMSA) that enhances the model’s capacity to capture detailed features of low-contrast cracks. To validate the effectiveness and superiority of the AnisoNet, multiple experiments were conducted on several public datasets. The results show that the proposed method outperforms existing state-of-the-art models in crack detection accuracy, achieving improvements of 6.1% in mAP@50 and 6.8% in mAP@95 compared to baseline models, while AnisoNet exhibits excellent generalization capabilities across various complex scenarios, including different acquisition viewpoints, challenging illumination conditions, and background interferences. Furthermore, to validate the method’s engineering feasibility within real-world ITS scenarios, a real-time detection system was implemented on a UAV-based edge computing platform. Field tests confirm that AnisoNet achieves an inference speed of approximately 26 FPS on embedded devices, successfully substantiating its applicability for low-latency pavement health monitoring tasks.