Low-level features in present railway fastener detection algorithms have poor descriptive ability and high feature dimension. Aiming at these problems

a fastener detection model which fuses potential semantic topics based on two kinds of low features is proposed. Latent Dirichlet allocation（ LDA） model is designed to capture latent semantic topics of histogram of gradient（ HOG） feature and local binary pattern（ LBP） feature fastener pictures. Topic distribution of each feature is fused by different weight

which is determined by the LBP feature graphs and the gradient images. Vectors is used to train classifier and judge state of tested fasteners. Experimental results show that the proposed algorithm has lower loss and false detection rate and stronger detectability than existing detection methods.