📄 中文摘要
BEiT v2 通过训练机器预测紧凑且有意义的图像片段，即视觉标记，提升了对图像的理解能力。这种方法不仅填补了照片中的缺失部分，还能推测出缺失部分的含义，从而使模型更好地识别物体和场景。与传统的像素修补方法不同，BEiT v2 学会想象缺失部分的内容，这增强了系统的整体场景感知和上下文理解能力。通过将图像分块组合，模型在分类照片和勾勒物体方面表现出更强的能力。实验结果表明，该方法在大型图像测试中达到了更高的准确率，超越了以往的技术。
BEiT v2：使用向量量化视觉标记器的掩蔽图像建模
出处: BEiT v2: Masked Image Modeling with Vector-Quantized Visual Tokenizers
发布: 2026年2月15日
📄 English Summary
BEiT v2: Masked Image Modeling with Vector-Quantized Visual Tokenizers
BEiT v2 enhances machine understanding of images by training it to predict compact and meaningful segments known as visual tokens. This approach not only fills in missing parts of a photo but also infers their meanings, allowing the model to better recognize objects and scenes. Unlike traditional pixel patching methods, BEiT v2 learns to imagine what should be there, which strengthens the system's overall scene perception and contextual understanding. By grouping image patches, the model exhibits improved capabilities in tasks such as photo classification and object outlining. Experimental results demonstrate that this method achieves higher accuracy in large image tests, surpassing previous techniques.