Abstract Low signal-to-noise ratio (SNR) in wideband radar reconnaissance degrades performance of existing signal type recognition methods. To achieve higher correct recognition rate in low SNR levels, this paper proposes a recognition approach containing feature extraction and automatic estimation for signal type number. First, feature vectors of radar pulses are constructed from maximum energy angle (MEA) and maximum energy slice (MES) of ambiguity function (AF). To improve applicability in practical implementing, AF grids method is developed to obtain MEA and MES. A novel algorithm for denoising MES with geometric shape constraint is also proposed to improve validity of features with low SNR. Second, for realistic reconnaissance lacking prior knowledge, a type number estimation method is developed that costs little extra processing burden. Simulation results indicate that the proposed denoising algorithm brings high stability and robustness to extracted features in low SNR levels. The recognition approach in this paper is applicable to more realistic scenarios and achieves higher correct recognition rate with respect to low SNR compared with existing methods.