Recently,　in　VLSI　design　for　manufacturability　(DFM),　capturing　and　representing　the　intrinsic　characteristics　of　a　layout　is　of　great　importance.　Especially,　there　has　been　revival　of　interest　in　applying　machine　learning　techniques　into　DFM　field.　Feature　extraction　of　layout　patterns　is　imperative　before　feeding　into　learning　models　so　that　feature　representation　directly　affects　performance　of　machine　learning　model.　In　this　paper,　a　literature　review　of　recent　progress　on　VLSI　layout　feature　extraction　is　firstly　conducted.　Then,　for　the　first　time,　we　propose　a　dictionary　learning　approach　wrapped　in　an　online　learning　model　in　applications　of　VLSI　layout　such　as　sub-resolution　assist　feature　(SRAF)　generation　and　hotspot　detection.　With　mapping　original　features　into　a　sparse　and　low-dimension　space,　dictionary　learning　model　is　benefit　to　calibrate　a　machine　learning　model.　The　experimental　results　show　that　our　method　not　only　improves　the　accuracy　of　hotspot　detection　but　also　boosts　F1　score　in　machine　learning　model-based　SRAF　generation　with　less　time　overhead.