Recognition of radar emitter signals is one of core elements in radar reconnaissance systems. In order to attain a higher correct recognition rate of radar emitter signals under condition of low signaltonoise (SNR) ratio, a novel method based on Main Ridge Slice of Ambiguity Function (MRSAF) and Deep Belief Network (DBN) is presented. Firstly, the singular value decomposition (SVD) is preprocessed for noise reduction, and then this paper calculates the AF of the sorted signal and ascertains the main ridge slice envelope. To improve the recognition performance, the SVD is employed to eliminate the influence of noise on the main ridge slice envelope. A DBN model is established on the stacked Restricted Boltzmann Machines (RBM) and the labeled data with the supervisory finetuning model parameters are used to complete the training. Finally, the model is used to achieve the radar emitter signals recognition and classification. The simulation results indicate that the novel algorithm provides significant performance and the validity and application value of this algorithm are verified. Compared to the existing methods, the novel method can achieve a higher correct recognition rate even at a low SNR.