Abstract: Aiming at the problem of cumbersome manual parameter adjustment and large model capacity of rail surface defect detection model, a compression method of rail surface defect detection model incorporating Bayesian optimization is proposed.Firstly, the Bayesian optimization algorithm is used to obtain the optimal hyperparameters that balance the sparsity and accuracy of the model; secondly, the obtained hyperparameters are introduced into the sparse training and the detection model is compressed under two strategies focusing on the compression ratio and the accuracy; finally, the parameters of the compressed model are corrected through the knowledge distillation to compensate for the loss of accuracy caused by the sparse training.The experimental results show that the compression ratio of the lightweight track surface detection model obtained by this method under the two sparse training strategies can reach 96.35% and 93.22%,respectively, and the detection speed is improved by more than two times after the hardware is deployed, which can avoid the negative impact of manual parameterization on the compression accuracy.