With　the　application　of　artificial　intelligence　technology,　face　detection　is　now　not　only　concerned　with　accuracy　but　detection　speed　as　well.　However,　most　previous　works　have　relied　on　heavy　backbone　networks　and　required　prohibitive　run-time　resources,　which　seriously　restricts　their　scope　for　deployment　and　has　resulted　in　poor　scalability.　In　this　study,　we　used　YOLOv5s,　which　has　a　good　detection　rate　and　accuracy,　as　the　baseline　network.　First,　we　added　a　none-parameter　channel　attention　self-enhancement　module　to　allow　the　backbone　of　the　network　to　capture　the　characteristic　features　of　the　face　more　effectively.　Second,　a　low-level　feature　fusion　module　was　added　to　enhance　the　features　of　shallow　neural　layers　and　then　fuse　them　with　the　features　of　deeper　layers.　Third,　a　receptive　field　matching　module　allows　the　network’s　perceptual　field　to　better　match　the　scale　of　actual　faces.　Finally,　contextual　information　based　on　face　key　points　allows　the　face　detector　to　exclude　more　cases　of　error　and　missed　detections.　On　the　most　popular　and　challenging　face　detection　dataset,　WIDER　FACE,　our　model　performed　better　than　the　original　network,　with　improvements　of　3.8,　4.4,　and　11.6%　on　the　easy,　medium,　and　hard　subsets,　respectively,　and　achieved　a　rate　higher　than　72　FPS,　which　meets　the　real-time　requirements.　©　2022,　The　Author(s),　under　exclusive　licence　to　Springer-Verlag　London　Ltd.,　part　of　Springer　Nature.