Surface electromyogram (sEMG) signals are physiological signals that are widely applied in certain fields. However, sEMG signals are frequently corrupted by noise, which can lead to catastrophic consequences. A novel scheme based on complementary ensemble empirical mode decomposition (CEEMD), improved interval thresholding (IT), and component correlation analysis is developed in this study to reduce noise contamination. To solve the problem of losing desired information from sEMG, an sEMG signal is first decomposed using CEEMD to obtain intrinsic mode functions (IMFs). Subsequently, IMFs are selected via component correlation analysis, which is a measure used to select relevant modes. Thus, each selected IMF is modified through improved IT. Finally, the sEMG signal is reconstructed using the processed and residual IMFs. Root-mean-square error (RMSE) and signal-to-noise ratio (SNR) are introduced as evaluation criteria for the sEMG signal from the standard database. With SNR varying from 1 dB to 25 dB, the proposed method increases SNR by at least 1 dB and reduces RMSE compared with stationary wavelet transform and other denoising algorithms based on empirical mode decomposition. Moreover, the proposed method is applied to hand motion recognition. Results show that the rate of the denoised sEMG signal is higher than that of the raw sEMG signal.