Deep　supervised　hashing　is　popular　for　large　scale　image　retrieval　in　recent　years.　Most　existing　deep　hashing　methods　which　preserved　the　relationship　between　Hamming　distance　and　inner　product　to　implement　the　distance　metric　of　the　binary　codes.　And　they　usually　used　sign　function　made　the　back　propagation　impractical.　In　this　paper,　we　propose　a　novel　deep　supervised　hashing　method　called　discrete　classification　optimization　hashing　(DCOH).　A　new　loss　function　is　designed　by　introducing　the　constraints　of　cosine　similarity　and　classification　label.　It　can　increase　the　similarity　between　the　original　image　and　the　corresponding　binary　codes.　Furthermore,　the　sigmoid　function,　which　can　iteratively　approach　the　sign　function,　is　used　to　resolve　the　problem　of　the　back　propagation　impractical.　The　experimental　results　confirm　the　effectiveness　of　the　proposed　algorithm.　©　2019　IEEE.