The purity of reference data affects the detection capability of DTMB passive radar, and the accuracy of reference channel estimation is a key factor to affect the recovery of reference data. In view of this problem, the channel estimation method based on the deep learning theory is introduced into the passive radar reference channel estimation process. The reference channels are modeled by the autoregressive model and the reference channel estimation network is built. After iterative training, the reference channel response is estimated. Compared with the traditional algorithm, the estimation accuracy of reference channel based on deep learning is improved effectively, and the detection performance of radar is improved.