In　order　to　improve　the　segmentation　accuracy　of　graph＇s　minimum　spanning　tree　and　reserve　more　edge　details,　a　new　image　segmentation　method,　which　is　on　the　basis　of　non-subsampled　Contourlet　transform　(NSCT)　and　improved　graph＇s　minimum　spanning　tree　(MST)　is　proposed.　Firstly,　an　image　is　decomposed　into　a　low-frequency　sub-band　and　several　high-frequency　direction　sub-bands　through　NSCT　decomposition.　Secondly,　the　high-frequency　direction　sub-bands　are　denoised　according　to　the　improved　Bayes　shrink　threshold,　and　edge　points　are　detected　according　to　the　module　maxima.　Then,　a　multi-scale　multi-direction　MST　edge　weight　is　constructed　according　to　the　grey　value　of　low-frequency　sub-band　and　the　coefficients　of　high-frequency　sub-bands,　and　the　edge　weight　of　edge　points　is　increased.　Moreover,　MST　algorithm　is　improved　in　two　main　aspects,　one　is　the　function　of　intra-regional　and　inter-regional　differences,　and　the　other　is　the　re-merge　mechanism　after　segmentation.　Thus,　the　impact　of　noises　or　isolated　points　can　be　reduced.　Finally,　the　optimal　position　adjustment　strategy　of　harmony　search　is　improved　and　adopted　to　find　the　optimal　parameters　of　global　optimal　MST　segmentation　results　adaptively.　Experimental　results　show　that,　in　comparison　with　other　improved　MST　algorithms,　the　proposed　method　improves　both　anti-noise　performance　and　segmentation　accuracy,　and　helps　obtain　images　with　higher　segmentation　accuracy　and　better　edge　details.　©　2017,　Editorial　Department,　Journal　of　South　China　University　of　Technology.　All　right　reserved.