Machine　learning　has　provided　many　good　tools　for　superresolution,　whereas　existing　methods　still　need　to　be　improved　in　many　aspects.　On　one　hand,　the　memory　and　time　cost　should　be　reduced.　On　the　other　hand,　the　step　edges　of　the　results　obtained　by　the　existing　methods　are　not　clear　enough.　We　do　the　following　work.　First,　we　propose　a　method　to　extract　the　midfrequency　features　for　dictionary　learning.　This　method　brings　the　benefit　of　a　reduction　of　the　memory　and　time　complexity　without　sacrificing　the　performance.　Second,　we　propose　a　detailed　wiping-off　total　variation　(DWO-TV)　regularization　model　to　reconstruct　the　sharp　step　edges.　This　model　adds　a　novel　constraint　on　the　downsampling　version　of　the　high-resolution　image　to　wipe　off　the　details　and　artifacts　and　sharpen　the　step　edges.　Finally,　step　edges　produced　by　the　DWO-TV　regularization　and　the　details　provided　by　learning　are　fused.　Experimental　results　show　that　the　proposed　method　offers　a　desirable　compromise　between　low　time　and　memory　cost　and　the　reconstruction　quality.　(C)　2015　SPIE　and　IS&T