The complex radar emitter signal recognition methods have problems of poor anti-noise performance, low recognition rate, etc. To address these issues, we propose a new recognition method based on ambiguity function contour lines and stacked denoising autoencoders. First, the ambiguity function is processed by the Gaussian smoothing and the contour lines are calculated by linear interpolation. Then, principal component analysis is used to reduce its feature dimension. The main ambiguity energy information is remained. Finally, deep learning stacked denoising auto-encoders are established to learn and extract the deep and more ubiquitous features of contour lines. The Softmax classifier is used to classify them. Simulation experiments show that the overall average recognition rates of six types of typical radar signals are all above 99. 83% when the signal-noise ratio is 0 dB. The recognition rate can also reach 83. 67% when the signal-noise ratio is -6 dB. Results prove that this method has good performance and feasibility under the extremely low signal-noise ratio conditions.