With　the　Internet　technology　development　and　the　popularization　of　networks,　multimedia　information,　especially　images　and　visual　information　because　of　its　rich　and　varied　information,　has　become　an　important　part　of　information　retrieval,　in　which　video　information　which　is　the　image　information　as　a　basis.　In　the　image　retrieval　technology,　in　order　to　retrieve　the　results　closer　to　people＇s　thinking,　how　to　use　semantic-based　content-based　image　retrieval　using　low-level　features　of　the　image　fit　the　human　to　become　a　key　high-level　semantics.　Support　Vector　Machine　is　considered　a　highly　effective　performance　of　the　classifier　are　superior　in　many　occasions　can　be　applied.　On　this　basis,　direct　push　support　vector　machine　(TSVM)　is　a　combination　of　support　vector　machine　algorithm,　to　achieve　an　efficient　classification　algorithm.　This　paper　puts　forward　a　new　algorithm　in　the　original　TSVM　algorithm　based　on　the　use　of　labeled　samples　at　the　same　time,　consider　a　sample　of　the　classifier　without　labeling　effects,　and　add　in　the　original　part　of　the　screening　process　conditions,　as　well　as　process　improvement,　making　the　new　algorithm　in　time　complexity　have　significantly　decreased,　while　no　significant　effect　on　algorithm　results.　©　2010　IEEE.