Aiming at the problem that the target information is difficult to effectively highlight due to the poor reconstruction effect of ghost imaging, combined with the advantages of autoencoder neural network in noise reduction, a ghost imaging optimization method is proposed. In this method, a handwritten digital data set is used as a sample, and a noise reduction network model is designed based on the second-order correlation of the detection data to obtain the initial ghost image. The network model uses the Leaky relu linear activation function to solve the problem of network oversaturation and cell death, and the effectiveness of the proposed network model is verified through 10,000 testing sample sets. By comparing the quality of ghost images before and after optimization at different sampling rates, the analysis results show that the peak signal-to-noise ratio of ghost images after optimization is increased by 87.02%/93.99%, 81.97%/85.90%, 27.22%/18.16%, respectively; at the same time, the contrast is increased by 479.03%/363.79%, 380.42%/272.91%, 38.76%/31.05%, respectively, compared with CGI, DGI, and CS respectively.