Aimed at the problems that the radar waveform multicriteria optimization target function is difficult to establish, and in order to reduce the uncertainty of target response and to improve the radar detection performance, a radar waveform design method based on deep neural network is proposed. First, deep neural networks (DNNs) are designed according to the radar echo data. Then, the signals generated based on the SignaltoNoise Ratio (SNR) and the Mutual Information (MI) criteria are randomly mixed, and the corresponding environmental information is used to form a training set, and the DNNs are trained. Finally, taking another part of the signals generated based on the mutual information criterion and its corresponding environmental information as test sets, this paper utilizes the DNNs for generating signals and testing. The experimental results show that if the signals generated by the method taken as a radar emission waveform compares to the signals generated based on the MI criterion alone taken as a radar transmission waveform, the mutual information of the radar echo and the target is increased by 21.37 nat, and the signals of the radar receiving signals are improved. The noise ratio is increased by a maximum of 1.35 dB. Compared with the chirp signals, the corresponding mutual information is increased by 950.76 nat, and the corresponding signaltonoise ratio is increased by 18.23 dB.