Fingerprint　recognition　is　regarded　as　the　most　reliable　and　accurate　biometric　identification　system.　Most　traditional　fingerprint　identification　systems　use　digital　image　processing　technology　rather　than　new　deep　learning　technology.　While,　provided　that　the　image　quality　is　bad,　the　accuracy　of　the　traditional　fingerprint　system　maybe　could　not　reach　the　ideal　value.　In　this　paper,　we　created　three　different　Siamese　networks　based　on　CNN　structures　to　compare　the　impact　of　network　structure　on　fingerprint　recognition　accuracy.　The　three　CNN　networks　are　self-defined　net,　AlexNet,　and　VGG.　We　use　dynamic　adjustment　of　the　threshold　to　make　it　more　adaptable　to　a　variety　of　data　sets.　The　biggest　difference　between　these　three　networks　is　the　number　of　parameters.　In　the　experiment,　we　use　the　data　set　of　1000　factorial　pairs,　and　the　average　accuracy　of　fingerprint　recognition　is　97%,　94%,　and　51%　respectively.　And　then,　we　mainly　discuss　that　the　Siamese　network　based　on　CNN　has　an　excellent　advantage　in　fingerprint　recognition,　and　the　struct　of　CNN　is　the　key　to　affect　the　accuracy　of　fingerprint　recognition.　©　2021　IEEE.