Aiming at the low accuracy of existing night image visibility detection algorithms, this paper proposes a night visibility classification algorithm based on stable light sources. Firstly, all stable light source street lights in the image are detected by the target detection network, and all light source blocks in the image are obtained. Secondly, the light source block is classified by the classification network. At the same time, all the light source blocks are sorted by size and average size and the corresponding weights are obtained. Finally, the classification results and weights of light source blocks are combined with statistical analysis to classify the visibility level of the night image. The experimental results show that the accuracy of the night visibility classification algorithm in this paper reaches 77.6% in the real social data set, and the classification results are more accurate than the existing methods, and have good generalization in different scenarios.