StyleBiFormer:融合BiFormer和可逆残差块的风格迁移

打开文本图片集

中图分类号：TN911.73-34；TP391 文献标识码：A 文章编号：1004-373X（2025）15-0037-08

StyleBiFormer：Fusing BiFormer and reversible residual blocks for style transfer

LIUZhaoyang，FANYao，JIANGMin （College of Information Engineering，Xizang Minzu University，Xianyang 712O82， China）

Abstract：The StyleBiFormer，a style transfer model thatincorporatesBiFormerand reversibleresidual blocks （RRBs）is proposed toeliminatethe limitationsofCNNincapturingglobal featuresofanimageandtheshortcomingsof the Transformerin parameterutilization.Firstly，thesemanticpositionalencoding（SPE）isusedtoprovidethemodelwithsemanticandpositional informationofPatch.Then，theglobalandlocalfeaturesoftheimageareextractedwiththeencodersandreversibleresidual blocks，whichavoidsthefeatureinformationlosseffectively.Finall，adecoderisusedtofusethemostmatchedcontentfeatures andstylefeatures togenerateamore visuallappealing image.Qualitativeandquantitativeexperimentsprovedtheeffectiveness ofStyleBiFormer.Incomparisonwiththoseoftheothermainstreammethodswithbestefectiveness，theSIM（structural similarity）ofthemodelincreases by 7% and its Gram Loss decreases by 3% .Inaddition，thevisual effect ofthegenerated images isbeter，avoiding thefeature leakagewhilepreserving theglobalfeatures.Theproposedmodelcanefectivelyextractand fusethecontentfeaturesandstylefeatureswiththehighestsemanticsimilarity，andachieveavividstyletransferefect，othat the original imagecan presenta brand-new stylewhileretaining theoriginal semantics.Tosumup，theproposed modelcan meetthe task of image style transfer inreality.

Keywords：CNN;style transfer;BiFormer;RRB; SPE;semantic similarity

0 引言

图像风格迁移（ImageStyle Transfer）是一种将计算机视觉（ComputerVision）应用于艺术创作的技术。（剩余12280字）