In　order　to　improve　the　effect　of　CNN　feature　driven　flotation　performance　recognition　under　small-scale　training　set,　a　method　of　flotation　performance　recognition　based　on　adaptive　transfer　learning　and　CNN　features　extraction　of　foam　infrared　and　visible　images　is　proposed.　Firstly,　a　dual-modality　CNN　feature　extraction　and　recognition　model　based　on　AlexNet　was　constructed,　and　the　structural　parameters　of　the　model　were　pre-trained　through　RGB-D　large-scale　data　set.　Secondly,　a　series　of　double　hidden　layer　automatic　encoder　extreme　learning　machine　is　used　to　replace　the　full　connection　layer　of　the　pre-training　model,　so　that　the　dual-modality　CNN　features　can　be　fused　and　abstracted　layer　by　layer,　and　then　the　decision　is　made　by　mapping　to　higher　dimensional　space　through　the　kernel　extreme　learning　machine.　Finally,　the　floatation　small-scale　data　set　is　constructed　to　train　the　migrated　model,　and　the　improved　quantum　wolf　pack　algorithm　is　used　for　model　parameter　optimization.　Experimental　results　show　that,　adaptive　transfer　learning　can　significantly　improve　the　accuracy　of　recognition　in　small　sample　data　sets,　the　accuracy　of　performance　recognition　using　dual-modality　CNN　transfer　learning　is　3.06%　higher　than　that　of　single-mode　CNN　transfer　learning,　and　the　average　recognition　accuracy　of　each　working　condition　reached　96.86%.　The　accuracy　and　stability　of　flotation　performance　recognition　is　greatly　improved　compared　with　the　existing　methods.　©　2020,　Science　Press.　All　right　reserved.