The　number　of　surface　defects　in　steel　has　always　been　a　key　measure　of　the　quality　of　steel　production.　Traditional　detection　methods,　such　as　the　strobe　method,　have　the　problems　of　slow　detection　reaction　and　low　accuracy.　The　object　detection　technology　based　on　deep　learning　can　effectively　improve　the　detection　performance　by　virtue　of　its　strong　real-time　performance　and　high　accuracy.　However,　in　the　actual　production　process,　complex　production　environments　and　other　industrial　factors　may　affect　the　efficiency　of　object　detection　technology.　We　used　an　improved　You　Only　Look　Once　(YOLO)　framework　to　identify　defects　on　the　steel　surface.　This　improved　version　of　the　YOLO　model　optimizes　the　original　architecture　to　enhance　the　detection　of　small　size　and　low　contrast　defects.　As　a　real-time　object　detection　system,　YOLO　is　able　to　predict　the　locations　and　category　probability　of　defects　directly　from　the　input　image　by　integrating　convolutional　neural　networks.　In　addition,　we　have　adjusted　the　loss　function　and　anchor　frame　strategy　of　the　model　to　improve　the　detection　accuracy　and　response　speed　in　complex　environments.　The　performance　of　the　model　in　different　types　of　defect　identification　was　analyzed,　and　the　applicability　and　efficiency　of　the　model　in　the　actual　steel　defect　dataset　NEU-DET　were　discussed.　©　2024　IEEE.