The lower signal-to-noise ratio of underground microseismic signals results in a decrease in signal picking accuracy. At present, signal denoising algorithms based on wavelet thresholding encounter problems such as poor generalization and difficulty in measuring thresholds when facing signals with low signal-to-noise ratios. To address this issue, this study investigates a fully supervised learning method based on complex wavelet transform for microseismic waveform denoising. The proposed method first utilizes a complex wavelet transform combined with a convolutional autoencoder to design an encoder-decoder with multiple convolutional and deconvolution operations to complete the image denoising. To verify the effectiveness of this method, Earthquake2023 is first constructed on Stanford’s Earthquake dataset for training and testing. It shows good fitting performance and training results. At the same time, a seismic phase picking method is designed based on the denoised signal obtained from this method and achieves high picking accuracy. This study designs multiple sets of comparative experiments, and the results show that the denoising method can effectively improve the peak signal-to-noise ratio and root mean square error of the signal, which have increased by 16 dB and 24% respectively. Moreover, the error of picking up P-wave and S-wave at the first arrival time is reduced by 0.3 ms compared to STA/LTA.