Abstract: Intelligently and efficiently classification of pavement cracks is a critical task for pavement management departments to make cost-effective repairing activities. However, most of the current methods, either manual or automated ones, have a certain dependence on human interventions. This fact significantly influenced the efficiency and the functionality of the automated systems in real applications. In this paper, an unsupervised deep learning-based solution for pavement crack classification was addressed, which fused the optimized AlexNet model and the adaptive faster peak density clustering (AFDPC) algorithm. Three aspects are considered in the optimization of the AlexNet model, which include the proposal of the adaptive learning rate, the introduction of the ReLU6_Swish activation function, and the employment of weight normalization. An AFPDC algorithm was optimized based on the realization of adaptive cutoff distance calculation and cluster center selection. The integrated pavement crack classification architecture is obtained by creating pseudo labels using the clustering results, which are used to train the optimized AlexNet. The fused unsupervised deep learning-based pavement crack classification model was trained on 3000 pavement crack images that include 850 transverse cracks, 845 longitudinal cracks, and 1305 alligator cracks. Meanwhile, another 1000 randomly selected crack images are used to test the performance of the model, which showed a mean average precision of 0.851. Moreover, the method was tested on four public image datasets of COCO, CUB, SUN, and AWA datasets, which were benchmarked against other advanced algorithms reported in recent research papers.