A　novel　technique　that　can　implement　face　detection　directly　in　the　wavelet　compressed　domain　is　presented　in　this　paper.　The　algorithm　takes　the　entropy　decoding　and　inverse　quantized　wavelet　transform　coefficients　of　JPEG2000　picture　as　input,　and　outputs　the　locations　of　the　detected　faces.　The　main　contribution　of　this　work　is　in　proposing　a　multi-level　gradient　energy　representation　of　face　pattern　based　on　wavelet　compressed　data,　which　permits　pertinent　high　contrast　facial　parts,　such　as　eyes,　nose　and　mouth,　to　be　highlighted　in　a　compact　mode　no　matter　the　face　is　big　or　small.　A　neural-network　based　classifier　is　designed　to　decide　a　gradient　energy　pattern　as　face　or　non-face.　In　contrast　to　the　traditional　spatial-domain　techniques,　the　proposed　compressed　domain　technique　eliminates　the　unnecessary　decompression　step　and　thus　has　lower　computational　complexity.　Moreover,　compared　with　the　previous　methods　based　on　DCT　compressed　domain,　the　proposed　multi-level　gradient　energy　presentation　removes　the　complex　spatial　scaling　operation　in　compressed　domain　and　overcomes　block　quantization　problem.　Based　on　test　results　of　a　variety　of　pictures,　the　presented　algorithm　was　found　to　be　more　efficient　and　effective　than　the　previous　related　methods.