In　recognizing　traditional　crops　seeds　like　maize　seeds,　we　usually　use　electrophoresis　assay　method,　fluorescence　scanning　method　and　chemical　assay　method.　These　methods　are　destructive　methods.　They　take　a　long　time　to　detect　and　are　demanding　of　professional　background　knowledge　and　hardware　conditions　etc.　What＇s　more,　these　methods,　based　on　BP　neural　network　and　support　vector　machine　(SVM)while　taking　a　long　time　to　detect　are　less　accurate　in　process　of　classification.　In　this　paper,　based　on　the　computer　vision　technology,　we　proposed　a　new　method　for　the　classification　of　maize　seeds,　a　method　based　on　multi-scale　feature　fusion　and　extreme　learning　machine.　First,　we　extract　the　multi-scale　fusion　feature　of　maize　seeds.　Second,　based　on　extreme　learning　machine,　we　construct　the　classifier　model　of　maize　seed.　Third,　because　of　the　window　of　image　in　the　case　of　multi-scale　detection　has　the　problem　of　capturing　the　same　object　seed　with　many　overlapping　windows,　we　put　forward　a　kind　of　window　fusion　algorithm　to　solve　it.　The　simulation　results　show　that:　The　method　is　able　to　identify　the　maize　seeds　accurately.　Using　this　method　the　accuracy　of　classification　of　maize　seeds　can　reached　97.66%　and　the　error　rate　is　less　than　0.1%.　Compared　with　the　traditional　methods,　the　method　we　proposed　can　improve　the　speed　of　detection　and　the　accuracy　of　classification,　and　has　no　strict　hardware　requirements.