結(jié)合多尺度與多層級聚合的卷軸畫圖像描述模型

打開文本圖片集

DOI：10.16652/j.issn.1004-373x.2025.17.007引用格式：，，.結(jié)合多尺度與多層級聚合的卷軸畫圖像描述模型[J].現(xiàn)代電子技術(shù)，2025，48（17）：41-47.

關(guān)鍵詞：圖像描述；卷軸畫圖像；多尺度特征；非對稱卷積；多層級聚合解碼；Transformer中圖分類號：TN919-34；TP391 文獻(xiàn)標(biāo)識碼：A 文章編號：1004-373X（2025）17-0041-07

Scroll painting image caption model combining multi-scale and multi-level aggregation

YUE Chaoyang1，HU Wenjin12， ZHANG Fujun1 （1.KeyLabotofigsticduraloutinstrutiostnUesityou; 2.SchoolofMathematicsandComputerScience，NorthwestMinzu University，Lanzhou73oo3o，China）

Abstract：Thecrollpaintingimageshavediferentsizesandacertainspatialdistributioncharacteristics，andthetrasforer based encodinglayerisprone tolosingkeyimage information，soascroll painting imagecaptionmodel MMAcombiningmultiscaleandmulti-levelaggregationisproposed.Inthestageofencoding，byintroducingasymmetricconvolutionandmulti-scale featuremodules，theabilityof theconvolutionlayertoobtainspatialinformationcanbeimprovedefectivelyandtheglobaland localmulti-scalecontextualinformationofthescrollpaintingimagecanbeintegratedtoobtainafeaturerepresentationwithrich semanticinforatio.Amulti-evelaggationnetworkisdsigndintestageofeoding.Byggatingtefeatursofit codinglayers，thesemanticinformationofthehigh-levelcodinglayerandthecontentinformationofthelow-levelcodinglayeare utilizedeffctivelytherebyaleviatingtheinformationlossefetivelyExperimentalresultsshowthatthemodelachievsood resultsonthescrollpintingdataset，improving BLEU-4andMETEORby26.7%and0.9%，respectively，incomparisonwiththe NIC （neural image caption） model，and generates more accurate description sentences.

Keywords：imagecaption;scroll paintingimage；multi-scale feature；asymmetricconvolution；multi-levelaggregation decoding; Transformer

0 引言

卷軸畫作為我國獨特的繪畫藝術(shù)形式，通過對卷軸畫進(jìn)行圖像描述能夠幫助人們賞析和理解卷軸畫，促進(jìn)文化交流，同時為卷軸畫的研究和數(shù)字化保護(hù)提供技術(shù)支持。（剩余10500字）