Learning-based　image　super-resolution　is　one　of　the　most　promising　approaches　to　solve　the　image　super-resolution　problem.　A　novel　pre-classified　learning　based　image　super-resolution　algorithm　is　proposed　to　reduce　the　complexity　of　full　searching　and　to　avoid　mismatching.　A　texture-based　pre-classified　process　is　used　to　select　a　subset　of　samples.　Then,　the　best-matching　samples　are　searched　among　the　selected　subsets.　In　the　proposed　algorithm,　the　complexity　of　the　searching　process　is　effectively　reduced　by　the　texture-based　pre-classified　process.　Furthermore,　using　the　texture　features,　the　mismatching　probability　is　reduced.　Experimental　results　show　that　both　the　visual　quality　and　the　run-time　are　improved.