Denoising, noise or interferences are removed from recorded signal to enhance the signal-to-noise ratio (SNR), is a crucial and ubiquitous step in the procedure of signal processing, especially for neurophysiological signal. This step facilitates following processing, such as feature extraction, classification, and data analyses. Conventional methods are based on the principle of separating noise components from the recorded signal and removing them, but these methods do not remove noise completely. In particular, conventional methods seems powerless to eliminate irregular and occasional noise bursts, which are caused by transient electrode contacting problem, head movements, or unpredictable factors. In this paper, we tackled the problem of noise removal from a new perspective, which is opposite to the conventional methods. Data portions that are contaminated by noise are entirely removed and then restored according to their relationships with the remaining signal. The rationale of this procedure is to purify the signal through addition rather than deduction that is normally executed in conventional methods. The results of both synthetic data and real EEG demonstrated that our idea is feasible and provides a new promising manner for noise removal.