视觉-语言模型研究综述OA

In recent years,with the rapid development of multimodal learning,Visual-Language Models(VLMs)have demonstrated significant performance advantages in cross-modal tasks such as image captioning and visual question answering.By combining visual and linguistic information and leveraging large-scale image-text pairs available from the internet for pretraining,VLMs have become a research hotspot in the field.However,systematic reviews of VLMs,especially those including performance comparisons,analysis,and comprehensive reviews of end-to-end training processes,remain scarce.Therefore,we provide a comprehensive overview of the latest advancements in VLMs as of 2025,covering:classification and discussion of original text and image feature processing methods;classification and review of mainstream modal interaction strategies;review and discussion of classic and cutting-edge model architectures;a systematic summary of popular VLMs;benchmarking and discussion of current transfer learning methods in terms of per-formance,and domain generalization.The three future research directions are proposed.