CoCrack: A completeness-oriented framework toward effective and efficient pavement crack detection

doi:10.1016/j.aei.2026.104915

CoCrack: A completeness-oriented framework toward effective and efficient pavement crack detection

2026-06-01

0

PRE

AI

DOI：10.1016/j.aei.2026.104915

Original

Original request for help

Save

Abstract

En 中文

Timely pavement crack detection is crucial for extending road service life and ensuring traffic safety. However, in complex scenarios, existing methods struggle to balance effectiveness and efficiency, often yielding incomplete segmentation results such as fragmentation and missed detections-especially for minor cracks or under strong background interference. To address this, we propose an effective and efficient pavement crack detection framework named CoCrack. It formalizes the engineering requirements related to completeness and deployability through the following targeted innovations, while concurrently strengthening fine-grained feature capture and global context modeling capabilities. First, a novel reparameterized gating inverted-bottleneck module serves as the core encoding-decoding component: it enables multi-branch representation and gating screening of fine-grained crack features during training, and is reparameterized into a single-branch equivalent structure during inference, thus ensuring inference efficiency while expressively capturing local details and minor branches of cracks. Second, a novel scale-incremental large kernel aggregation network acts as the main feature extractor: it assembles large depthwise convolution kernels with sizes following an arithmetic sequence to fuse cross-span crack features, and leverages forward-reverse attention synergy to enhance spatial discriminability, significantly expanding the receptive field for full exploitation of crack global contexts. Third, a concise asymmetric U-shaped architecture with Transformer-style patch operations facilitates fast and low-wastage encoding-decoding, and a Lovasz-extended hybrid loss further suppresses missed detections. Extensive experiments on public datasets demonstrate that CoCrack (8.20 M parameters) and CoCrack-s (2.44 M parameters) significantly outperform state-of-the-art models across all segmentation metrics. Even the lightweight CoCrack-n (0.45 M parameters) exhibits highly competitive performance. Deployment tests on an in-vehicle edge device confirm CoCrack-n's effectiveness and efficiency under resource constraints. Overall, the proposed framework achieves superior performance and strong robustness, with segmented crack regions featuring ultra-high completeness and ultra-low missed detection rate. It holds great practical engineering value for pavement distress inspection and construction health monitoring.

Keywords:

Crack detection

Completeness-oriented segmentation

Lightweight model

Missed detection

Local details

Global contexts

CoCrack: A completeness-oriented framework toward effective and efficient pavement crack detection

Abstract

Journal

Organization

Cited Papers