Small-sample data-driven lightweight convolutional neural network for asphalt pavement defect identification

Automatic detection technology provides a reliable method for civil engineering distress detection. However, to overcome limitations of computational resources and the significant cost of image acquisition, this study proposes a simplified network parameter-based pavement crack classification network (PCCNet) to achieve efficient and robust crack classification. Firstly, a lightweight classification model is developed based on a shuffle unit and inverted residual architecture, designed to deliver high-performance pavement crack classification with limited computing resources. Secondly, a novel training method is proposed to accurately identify pavement defects on small-sample pavement images datasets. Additionally, the interpretability of neural network in pavement defect detection is enhanced by visualizing training process. The results demonstrate that the model achieved a classification accuracy of 97.89 % on the augmented pavement image dataset and a classification accuracy of over 83 % on multi-source asphalt pavement images. Furthermore, visualizing intermediate features further enhanced the high-precision recognition ability of the lightweight model.