Towards Efficient and Explainable Road Crack Detection: A Transfer Learning-Driven SqueezeNet-based Approach

Abstract

Abstract Under long-term exposure to complex environmental conditions and variable traffic loads, road pavements are generally prone to typical early damages such as cracking. Such defects can lead to severe structural deterioration and pavement service life reduction if not detected and repaired in time. Traditional manual inspection of pavement defects is usually inefficient and highly subjective, making it difficult to balance accuracy and cost-effectiveness in large-scale assessments. In recent years, deep learning based road crack detection methods have received widespread attention. However, existing models often suffer from excessive complexity, high computational cost, and heavy reliance on large-scalelabeled datasets. To address these challenges, this study proposes an efficient and interpretable crack detection method that integrates SqueezeNet with transfer learning. First, a lightweight SqueezeNet architecture is employed to extract pavement crack features, which reduces parameter size and computation cost. Then, a transfer learning strategy is introduced to enhance model generalization and mitigate dependence on extensive annotation. Experiments were conducted using 25 000 pavement images collected from a primary roadway connecting Chengdu with surrounding cities in western China. Comparative analyses were performed against unsupervised learning methods, classical edge detection algorithms, and typical state-of-the-art deep learning models. Results indicate that the proposed model achieves high detection performane, specifically, accuracy (ACC), true positive rate (TPR), true negative rate (TNR), positive predictive value (PPV), and F1-score (F1) all exceed 91%. The error rate (ER) and false positive rate (FPR) remain below 7.5%. Moreover, gradient-weighted class activation mapping (Grad-CAM) visualization accurately localizes crack regions, elucidating the model’s decision-making process and enhancing interpretability. This study provides a high-accuracy, low-cost, and explainable solution for automated pavement crack detection, which has significant practical value for infrastructure health monitoring and maintenance decision-making processes.