In　recent　years,　Optical　Character　Recognition(OCR)　is　widely　used　in　machine　vision.　In　this　paper,　we　investigated　the　problem　of　optical　character　detection　and　recognition　for　Image-based　in　natural　scene.　The　Optical　Character　Recognition　is　divided　into　three　steps:　(1)　Selecting　the　candidate　regions　through　image　preprocessing.　(2)　The　detection　neural　network　is　used　to　classify　each　region.　The　purpose　is　to　retain　text　regions　and　remove　non-text　regions.　(3)　The　recognition　neural　network　is　used　to　identify　the　characters　in　the　text　regions.　We　propose　a　novel　algorithm.　It　integrates　image　preprocessing　with　Maximally　Stable　Extremal　Regions(MSER),　the　neural　network　architecture　of　detection　and　the　neural　network　architecture　of　recognition.　Compared　with　previous　works,　the　proposed　algorithm　has　three　distinctive　properties:　(1)　We　propose　a　new　process　of　OCR　algorithm.　(2)　The　application　scene　of　OCR　algorithm　is　the　images　of　natural　scene.　(3)　The　training　data　of　recognition　does　not　need　artificial　labels　and　can　be　generated　indefinitely.　Moreover,　the　algorithm　has　achieved　good　results　in　detection　and　recognition.