In　the　past　decade,　deep　learning　is　in　a　period　of　the　rapid　development,　widely　used　in　applications　on　different　fields.　In　general,　model　training　process　will　be　deployed　on　cloud　computing　side,　training　on　these　small　embedded　devices　is　not　recommended　since　it　has　the　lower-end　hardware　configuration.　Therefore　these　embedded　devices　are　usually　designed　for　inference.　In　this　paper,　a　new　image　recognition　framework　based　on　heterogeneous　multi　core　accelerator　was　established　to　achieve　deep　learning　prediction　process　and　improve　the　image　recognition　performance　of　embedded　devices.　At　firstly,　the　fundamental　principle　of　image　recognition　method　based　on　deep　learning　reviewed　as　the　basis　of　the　study.　And　secondly,　some　important　designs　of　CPU-Accelerator　heterogeneous　architecture　based　parallel　image　recognition　framework　included　data　splitting　strategy　framework　architecture,　data　structure　design　and　data　parallelism　were　proposed　to　improve　the　recognition　speed　and　the　computational　resource　efficiency.　Thirdly,　Xilinx　Zynq,　Adapteva　Epiphany　combined　hardware　platform　and　the　Rockchip　RK3288　hardware　platform　were　described　in　detail.　Finally,　an　experiment　of　handwritten　digits　recognition　was　conducted　to　evaluate　the　accuracy　and　performance　of　this　framework.　The　experimental　results　show　that　the　proposed　image　recognition　system　can　achieve　nearly　8　times　speedup　as　for　recognized　28x28　image　of　ten　handwritten　digits　and　nearly　60　times　speedup　as　for　recognized　32x32　image　of　ten　objects　classification　than　RK3288　board　which　has　the　newest　series　of　high-performance　Arm　core　CPU　as　the　control　included　4　Arm　A17　cores.　©　2018　ACM.