To enhance the accuracy and real-time performance of the intelligent assisted driving system in estimating the road adhesion coefficient, a deep learning algorithm based on visual information was developed for road recognition. The algorithm aims to achieve a pre-estimation of the road adhesion coefficient. A compression convolution mechanism was designed to reduce the network’s operation parameters. Additionally, the fully connection layer was replaced by the global average of the feature map to enhance the network’s fitting performance. Furthermore, a pavement recognition depth convolutional neural network called DW-VGG was constructed. The network was trained using a self-built pavement image dataset. The test results demonstrate that the DW-VGG network, utilizing the proposed multi-layer knowledge distillation algorithm, achieves a high recognition accuracy, with a classification performance evaluation index (F₁ score) of 96.57%. Moreover, it effectively reduces the network’s time and space costs, as it only takes 32.06 ms to identify a single image, and the prediction model size is merely 5.63 M.