There　are　a　lot　of　discontinuous　characteristics　such　as　grooves　and　the　steps　in　the　device　surface　with　the　groove　size.　The　least　squares　fitting　method　is　required　by　ISO　regulation　when　the　datum　is　extracted　for　its　quality　evaluation,　but　its　robustness　is　far　worse　than　filtering　method.　Fluctuations　which　are　called　the　pseudo　Gibbs　phenomenon　will　be　generated　in　the　discontinuous　surface　if　the　traditional　filtering　method　(such　as　Gaussian　filtering,　wavelet　filtering,　etc.)　is　used　directly.　A　new　filtering　algorithm　is　proposed.　By　gradient　processing　the　dual　tree　complex　wavelet　transform　(DT-CWT)　though　the　total　variation　minimization　theory　(TVM),　which　means　initializing　the　wavelet　coefficients　of　each　layer　by　using　the　threshold　value　method;　combining　the　idea　of　TVM　with　DT-CWT　with　projected　subgradient　method;　continuous　iteration;　and　finally　the　robustness　of　the　DT-CWT　transform　in　extracting　the　datum　which　has　plane　geometry　characteristic　is　enhanced.　The　simulation　results　show　that　compared　with　the　traditional　filtering　methods,　the　algorithm　has　the　optimality　in　the　characterization　of　precision　(maximum　reconstruction　error　and　the　mean　square　error),　and　its　assess　height　is　the　closest　to　the　ideal　height.　Assessment　carried　out　on　a　real　convex　sticker　block,　fluctuations　can　be　well　restrained　at　steps　whether　from　the　intuitive　3D　view　or　side　contrast　figure,　and　the　geometrical　characteristics　of　the　product　are　ensured　as　the　theoretical　height　deviation　with　the　actual　processed　standard　block　is　only　0.0403　μm,　and　an　effective　reference　extraction　method　is　provided　for　practical　engineering　application.　©　2015　Journal　of　Mechanical　Engineering.