In this paper, an innovative knowledge-based methodological framework to detect sleep slow waves (SSW) in the human sleep electroencephalogram (EEG) is proposed. Based on a restricted matching pursuit (RMP) algorithm, automatic pattern recognition of SSW is implemented using a small dictionary of Gabor functions modeling the target waveform morphological characteristics. The method describes EEG signals in terms of SSW properties and provides detection thresholds based on the largest MP coefficients. A computer software implementation of this new method was tested on a database of overnight polysomnographic recordings collected in 15 young healthy subjects and visually scored by a trained sleep expert. In addition to full automation and fast application, the results obtained from the RMP algorithm, and evaluated using a rigorous performance evaluation metrics, showed excellent agreement as compared with expert scoring, with 97% of correct detections and a concordance of 67% in SSW time position and duration. The performances demonstrated by this new method were superior to that of a canonical detection algorithm introduced earlier, with an equivalent sensitivity but a significant 12% increase in precision (p = 0.0002). By mimicking the way human processes information while scoring SSW, the RMP algorithm proves stable over time and sleep/wake states, and may thus be used with virtually no human intervention.