基于级联式逆残差网络的游戏图像多模态目标精准辨识研究

打开文本图片集

Research on game imagemultimodal objectprecise identification basedoncascadedinverseresidualnetwork

LIUJianzhi

（ShenyangLigongUniversity，Shenyang11oooo， China）

Abstract：Theobjects inthegame have complexfeaturessuch asshape，colorand texture.Inaddition，theobjects appear indiferentperspectives，scalesandpostures，allofwhichincreasethediicultyofbjectrecognition.Acascadediverse residualnetworkcanenhancethemultidimensionalfeaturesoftheobject，andcanrecognizetheobjectefectivelyeveninthe presenceofocclusion，deformation，etc.Therefore，agameimagemultimodalobjectpreciseidentificationmethod basedon cascadedinverseresidual network isproposed.Abackbonenetworkconsisting ofconvolutional layersandcascaded inverse residualmodules basedondepthwiseseparableconvolutiondesignisconstructed.Thisnetworkisused toextract thenputtd gameimage features preliminarily.Thechannelrearrngementisused toenhancetheinformation exchange among channels.The featureenancementnetworkisusedtoupsamplethefeaturemaps learnedbythebackbonenetwork.Themultimodalobject featuresareetractedincombinationwiththemulti-chanelfeaturefusion.Theobjectposition，direction，andotherinforation areoutputedbyapredictionnetwork thatcanachieveclasificationandregressontasks.Sofar，apreciseidentificationof multimodalobjects ingameimagesisachieved.Theexperimentalresultsshowthatthemethodcanachievetheidentificationof thecharacters，text，scene elements andother objects in the gameimages，with atraining lossofonlyabout O.O5andan F1 -score of 0.967. To sum up，the multimodal object recognition effect of game images is good.

Keywords：inverseresidual;；gameimage;；multimodal；channelrearrangement;SIoUlossfunction；objectidentification; convolutional layer;object location

0 引言

度不断提高，为玩家带来了更为真实和沉浸式的体验，但同时也给游戏图像目标辨识技术提出了更高的挑战。（剩余5474字）