K-means　clustering　separates　a　set　of　samples　into　several　groups　based　on　the　similarities　between　samples.　To　further　assess　the　nonlinear　correlation　between　high-dimensional　samples,　existing　deep　K-means　algorithms　just　exploit　an　auto-encoder　to　extract　the　inherent　features　of　samples,　and　then　perform　K-means　on　it.　From　this　letter,　we　present　the　real　deep　K-means　clustering　model　with　K　auto-encoders　where　K　is　the　number　of　clusters,　which　is　named　as　DKMA.　Specifically,　the　centroid　of　each　cluster　is　acted　by　one　auto-encoder,　rather　than　the　constant　vector　in　the　traditional　K-means.　Each　sample　decides　its　category　by　choosing　one　auto-encoder　which　reconstructs　the　sample　point　best.　The　extensive　experimental　results　indicate　that　the　our　present　approach　performs　better　than　the　other　clustering　algorithms.　©　2021　IEEE