平行图像:图像生成的一个新型理论框架

摘要

To build computer vision systems with good generalization ability, large-scale, diversified, and annotated image data are required for learning and evaluating the in-hand computer vision models.Since it is difficult to obtain satisfying image data from real scenes, a new theoretical framework for image generation is proposed, which is called parallel imaging.The core component of parallel imaging is various software-defined artificial imaging systems.Artificial imaging systems receive small-scale image data collected from real scenes, and then generate large amounts of artificial image data.In this paper, the realization methods of parallel imaging are summarized, including graphics rendering, image style transfer, generative models, etc.Furthermore, the characteristics of artificial images and actual images are analyzed and the domain adaptation strategies are discussed.%为了提高计算机视觉系统的泛化能力,要求利用大规模、多样化、带标注的图像数据集,对视觉模型进行充分的学习与评估.由于从实际场景中获取图像具有局限性,文中提出一种图像生成理论框架,称为平行图像.平行图像的核心单元是软件定义的人工图像系统.从实际场景中获取特定的图像"小数据",输入人工图像系统,生成大量新的人工图像数据.文中总结平行图像的实现方法,包括图形渲染、图像风格迁移、生成式模型等,并且对比分析人工图像和实际图像的特点,讨论领域适应策略.