The　paper　proposes　a　novel　super-resolution　reconstruction　algorithm　for　human　faces.　The　algorithm　extracts　training　examples　from　the　input　image　and　divides　them　into　several　classes　using　vector　quantization.　Then,　it　classifies　each　patch　from　a　low-resolution　image　as　one　of　these　classes.　Each　class　has　its　high-frequency　information　inferred　using　a　parallel　designed　multi-class　predictor,　which　is　trained　using　the　training　samples　from　the　same　class.　The　self-example　training　set　and　the　specific　domain　training　set　were　employed　in　investigation　of　the　impact　of　the　training　database.　The　experimental　results　showed　the　superior　performance　of　the　proposed　method　in　terms　of　both　the　reconstruction　quality　and　runtime.