基于预训练模型的小样本视觉问答研究综述OA

Traditional visual question answering(VQA)relies on large-scale annotated data to support model learning and reasoning,and thus performs poorly in low-data regimes.To address this limitation,few-shot visual question answering(FS-VQA)aims to achieve accurate understanding and multimodal reasoning under limited supervision,and has shown important application potential in low-resource scenarios such as medical imaging and remote sensing.However,challenges including cross-modal semantic bias,distribution mismatch,and insufficient generalization in open environments remain key bottlenecks for the development of FS-VQA.In this survey,this paper reviews the core theoretical connections between pre-trained models and FS-VQA,covering representative architectures of pre-trained models as well as funda-mental definitions of VQA and FS-VQA.It then categorizes and analyzes existing pretraining-based FS-VQA methods from three perspectives:supervision construction and distribution adaptation,cross-modal consistency and bias mitiga-tion,and open-knowledge enhancement and cross-task transfer.It summarizes the performance of representative models on mainstream benchmarks and cross-domain settings,and discusses their strengths,limitations,and emerging technical trends.Furthermore,the paper compiles typical applications of FS-VQA across different domains and outlines its transfer-ability and safety boundaries.Finally,it discusses current challenges in FS-VQA and highlights promising directions for future research.