Being　aimed　at　the　characteristic　in　complexity　and　levity　of　defect　image　surface,　a　novel　method　combined　NMI　feature　with　invariant　feature　in　time　domain　to　conceive　the　statistic　feature　of　defect　images　is　put　forward.　Simultaneously,　compactness　feature,　L-S　factor　feature　and　linearity　feature　in　the　rectangular　region　are　developed　as　one　basis　of　defect　classification.　Moreover,　in　frequency　domain,　a　method　which　can　extract　features　in　the　rectangular　region　of　central　bright　area　of　defect　spectrum　image　is　proposed,　and　maximum　difference　and　average　difference　of　gray　value　of　all　the　pixels　in　this　rectangular　region　are　developed　as　another　important　basis　of　defect　classification.　This　paper　also　applies　BP　neural　network　to　the　automatic　classification　of　defect　images,　constructs　the　defect　classifier　and　tests　six　types　of　common　defects　collected　from　online　data.　The　experimental　result　shows　that　the　new　features　extraction　method　increases　the　validity　of　classification　of　feature　greatly　and　this　BP　classifier　has　high　identification　accuracy　and　the　overall　recognition　rate　is　over　90%.This　new　technique　resolves　the　difficulty　of　defect　classification　on　defect　images　to　some　extent.