<p>Accurate pavement-type recognition remains a critical challenge for intelligent transportation systems. However, the general CNN-based methods, such as ResNet and VGG, exhibit significant limitations when addressing the high degree of similarity between surface modifications and dissimilar pavements caused by changes in illumination or partial shading. To address the challenges posed by complex texture variations and surface modifications across road pavements, in this study, an enhanced method for pavement-type identification is proposed. First, an application-oriented dataset encompassing seven pavement types is constructed based on existing open-source road surface classification datasets. Secondly, the EfficientNet-Global Attention Mechanism (GAM) model is developed through the integration of a GAM module into the EfficientNet architecture. Within this model, the GAM module undergoes a process where it synergistically refines channel–spatial features, utilizing 3D permutation and multilayer perceptron operations. This enables the effective isolation of discriminative patterns, such as crack density, from complex backgrounds. Then, to mitigate inter-class confusion, a label smoothing strategy is implemented, while cosine learning rate decay is employed to ensure stable convergence during training. The experimental results demonstrate that the proposed model achieves high-precision recognition of various pavement types, with an accuracy rate of 98.11%, while simultaneously maintaining computational efficiency. </p>