基于CLIP文本特征增強的剪紙圖像分類
打開文本圖片集
關(guān)鍵詞：視覺語言大模型；剪紙分類；小樣本分類；模態(tài)融合；提示學(xué)習(xí) 中圖分類號：TP391 文獻(xiàn)標(biāo)志碼：A 文章編號：1001-3695（2025）07-010-1994-09 doi：10.19734/j.issn.1001-3695.2024.11.0485
Abstract：Toaddressthechallengesoflarge modalitygaps between textand image featuresand insuficient classprototype representationin paper-cut image clasification，this paper proposed a CLIP-based textfeature enhancement method（CLIP visualtextenhancer，C-VTE）.Themethdextractedtext featuresthrough manualprompttemplates，designedavisual-textenhancement module，andemployedCrosssAtentionand proportionalresidualconnections tofuseimageandtextfeatures，therebyreducing modalitydiscrepancyandenhancing the expressiveabilityofcategoryfeatures.Experimentsonapaper-cutdataset andfourpublicdatasets includingCaltech01validatedits efectivenessForbase-classclasificationonthepaper-cutdataset， C-VTE achieved 72.51% average accuracy，outperforming existing methods by 3.14 percentage points. In few-shot classification tasks on public datasets，it attained 84.78% average accuracy with a 2.45 percentage-point improvement.Ablation experimentsdemonstratethatboth themodalityfusion moduleand proportional residual components contribute significantlytoperformanceimprovement.Themethodofersnovelinsightsforeficientadaptationof vision-languagemodelsindownstreamclassification tasks，particularly suited for few-shot learning and base-class dominated scenarios.
Key words：visual language large model；paper-cut classification；few-shotclasification；multimodal fusion；prompt learning
0 引言
在非遺領(lǐng)域中，剪紙主要是以圖片的形式存在，且種類復(fù)雜，數(shù)量繁多。（剩余22719字）