Near-infrared spectroscopy (NIRS) has been widely used to access the brain functional activity non-invasively, but traditional signal analysis methods are unsuitable for non-linear and non-stationary data. A novel signal processing method is developed by using Hilbert-Huang transform (HHT) to perform spectral analysis of functional NIRS signals sampled in a vision stimulation experiment based on event-related design. Compared with wavelet based multi-resolution analysis (MRA) and Fourier spectral analysis, we demonstrate that the HHT method is better in the extraction of task related signal for observation of activation in the prefrontal cortex. This method indicates a new analytical tool for functional NIRS signals and enables us to understand the episode of functional NIRS more precisely.