To　address　the　limitations　of　visible　image　feature-driven　flotation　performance　recognition　method,　a　new　flotation　performance　recognition　method　based　on　dual-modality　multiscale　images　CNN　features　and　adaptive　deep　autoencoder　kernel　extreme　learning　machine　was　proposed.First,　the　visible　and　infrared　images　of　foam　were　decomposed　by　nonsubsampled　shearlet　multiscale　transform,　and　a　two-channel　CNN　network　was　developed　to　extract　and　fuse　the　features　of　the　dual-modality　multiscale　images.Then,　the　CNN　features　were　abstracted　layer-by-layer　in　the　deep　learning　network,　which　was　connected　by　a　series　of　two　hidden　layer　autoencoder　extreme　learning　machine.Then,　the　decision　was　made　by　mapping　to　a　higher　dimensional　space　through　the　kernel　extreme　learning　machine.Finally,　the　quantum　bacterial　foraging　algorithm　was　improved　and　applied　to　optimize　the　recognition　model　parameters.　The　experimental　results　show　that　the　recognition　accuracy　using　dual-modality　multiscale　CNN　features　is　clearly　better　than　that　of　single　modality　multiscale　and　dual-modality　single　scale　CNN　features　at　a　confidence　level　of　2.65%.　Further,　the　adaptive　deep　autoencoder　kernel　extreme　learning　machine　has　better　classification　accuracy　and　generalization　performance.　The　average　recognition　accuracy　of　each　working　condition　reaches　95.98%.　The　accuracy　and　stability　of　flotation　performance　recognition　is　considerably　improved　compared　with　the　existing　methods.　©　2020,　Science　Press.　All　right　reserved.