Real-time　image　sensed　by　the　visual　sensor　usually　contains　a　lot　of　noise　information.　Model　reasoning,　and　pattern　recognition-oriented　CNNs　face　such　thorny　issues　as　excessive　computation,　poor　accuracy　and　high　resource　occupancy.　Hence,　CNN　architecture　was　heterogeneously　deployed　on　the　Zynq　platform　to　realize　hardware　acceleration　for　the　image　processing　algorithm.　MNIST　dataset　was　adopted　to　train　CNN　for　extracting　network　parameters　on　PC　terminal　under　the　Caffe　framework;　the　convolutional　layer　responsible　for　heavy　computational　load　was　deployed　onto　FPGA　for　parallel　computing　to　increase　system　speed;　input　layer　and　output　layer　responsible　for　a　small　amount　computation　were　placed　on　ARM　terminal　to　reduce　resource　consumption;　real-time　image　acquired　by　the　camera　was　binarized　to　highlight　image　features　and　improve　the　recognition　accuracy;　the　hardware　acceleration　performance　of　the　heterogeneously　deployed　CNN　was　verified　with　numerous　experiments　on　image　recognition　of　handwritten　numerals.　Experimental　results　indicated　that:　CNN　hardware　accelerator　kept　an　image　recognition　accuracy　up　to　99.02%　which　is　largely　equivalent　to　that　of　client　PC;　When　recognizing　a　single　piece　of　handwritten　numerical　sample,　under　the　use　of　optimized　instructions　and　100MHz　clock　frequency,　the　recognition　time　of　a　single　image　is　0.53s,　which　is　16　times　faster　than　pure　ARM　operation;　the　maximum　power　consumption　of　the　system　is　2.606W,　which　is　far　Lower　than　general-purpose　processors.　©　2021　IEEE.