Corner　point　is　the　pixel　with　a　high　curvature　on　image　edge.　It　is　a　key　feature　in　digital　image　processing.　Through　the　utilizing　of　corner　points　in　image　processing　tasks,　the　computational　complexity　can　be　highly　reduced.　This　paper　proposes　an　improved　corner　detection　algorithm.　A　technique　using　the　radius　of　the　fitting　circle　to　denote　local　curve　curvature　is　applied　on　the　basis　of　image　edge　after　Gaussian　smoothing,　and　then　a　method　using　threshold　is　provided　to　decide　the　support　region.　Finally,　mean　k-cosine　method　is　used　to　calculate　the　support　angle　and　the　false　corners　are　picked　out　from　the　candidate　corner　set.　Compared　with　classical　algorithm,　the　experimental　result　indicates　that　the　method　in　this　paper　is　efficient　and　accurate　when　extracting　corner　feature　from　2D　images.　©　2010　IEEE.