In　recent　years,　remote　sensing　satellites　have　developed　rapidly　and　accumulated　massive　high-resolution　image　data.　Using　deep　learning　to　solve　remote　sensing　image　object　detection　has　become　an　important　research　direction.　However,　due　to　shooting　from　a　bird＇s　eye　view,　the　object　has　an　oblique　Angle.　Because　of　the　perspective　problem,　remote　sensing　images　have　boundary　problem　when　the　object　is　near　horizontal.　When　there　are　a　large　number　of　densely　tilted　object　boxes,　using　horizontal　box　detection　method　will　cause　box　occlusion　thus　missing　or　mislabeling.　Therefore,　it　is　a　challenging　task　to　solve　remote　sensing　image　object　detection.　In　this　paper,　we　propose　the　network　structure　which　based　on　YOLO-V5　and　combined　with　CSL　method.　Moreover,　the　loss　function　is　improved　to　deal　with　the　overlapping　problem　of　tilted　object　boxes　effectively.　As　revealed　by　extensive　experiments,　the　average　accuracy　of　this　method　on　DOTA　data　sets　shows　a　competitive　advantage.　At　the　same　time,　the　size　of　the　network　is　only　16.3m,　which　is　obviously　better　than　the　comparison　method.　©　2022　IEEE.