An EfficientNet-global attention mechanism model based on road pavement type identification for vehicles

Abstract

This study presents an enhanced road pavement-type identification method for vehicles to address the challenge posed by the complex texture characteristics of various road surfaces. First, a comprehensive dataset, including seven pavement types, i.e., dry and wet asphalt, immersed asphalt, dry and wet concrete, and icy and snowy roads, was constructed based on the existing open-source road surface classification dataset and road testing experiments. Second, an EfficientNet-GAM model was developed, and a global attention mechanism (GAM) was incorporated to streamline the model’s complexity while preserving the global feature representation. Then, label smoothing was applied to the cross-entropy loss function, and a cosine learning rate decay strategy was employed to accelerate model convergence to improve the model’s learning of similar pavement information. Finally, the experimental results demonstrate the model’s effectiveness in identifying various pavement types with high accuracy (98.11%) and computational efficiency.