Benchmark Study on a Novel Online Dataset for Standard Evaluation of Deep Learning-based Pavement Cracks Classification Models

Highway agencies and practitioners expect to have the most efficient method with adequate accuracy when choosing a deep learning-based model for pavement crack classification. However, many works are implemented on their own dataset, making them hard to compare with each other, and also less persuasive and robust. Therefore, a Road Cracks Classification Dataset is proposed to serve as a standard and open-source dataset. Based on this dataset, a benchmark study of fourteen deep learning classification methods is evaluated. Two parameters, the Ratio of F1 and Training Time (RFT) and Ratio of F1 and Prediction Time (RFP), are proposed to quantify the efficiency of networks. The results show that ConvNeXt_base reaches the highest accuracy among all models but requires the longest training time. AlexNet takes the least training time among all models, but gains the lowest accuracy. Of the four crack types, the block crack has the lowest accuracy, which means it is the most difficult to detect. SqueezeNet1_0 has the highest efficiency among all models in converting the computing power to accuracy. Wide ResNet 50_2 consumes the longest prediction time among CNN models, while the ConvNeXt_base has the highest feasibility on real-time tasks. To implement a suitable deep learning-based pavement crack inspection, we recommend a good balance between computational cost and accuracy. Based on this, we provide practical recommendations according to different user groups.