【摘要】 人脸超分辨率能够有效提高低分辨率人脸图像的分辨率和质量，因而在视频监控、刑事侦查、娱乐等领域得到了广泛应用。然而实际场景中成像系统、记录设备、传输介质和处理方法不完善，导致噪声、模糊等降低图像质量的多种降质过程以不规则的方式组合，仅假设明确的降质过程来训练网络模型无法满足实际应用需求。针对这些实际场景中人脸超分辨率存在的多样化降质过程，分别介绍了非盲降质人脸超分辨率技术和盲降质超分辨率技术原理、人脸超分辨率领域常用的数据集和评价指标以及代表性工作的主客观重建结果。未来相关研究应聚焦多模态信息决策融合和张量融合，提升重建图像特征维度和时域相似性；通过大规模预训练和对抗学习等，提升模型泛化能力；研究身份一致性算法以及迁移学习等技术对处理复杂成像条件的影响。更多还原

【Abstract】 Face super-resolution enhances resolution and quality of low-resolution facial images,finding wide applications in fields of video surveillance,criminal investigation,and entertainment. However,in realworld scenarios,imperfect imaging systems,recording equipment,transmission media,and processing methods resulted in irregular combinations of degradation processes that reduce image quality,like noise and blurring. Training network models based on explicit degradation processes alone cannot meet practical needs.This paper reviews the principles of non-blind and blind face super-resolution techniques,the commonly used datasets and evaluation metrics and the subjective and objective reconstruction results of representative works in the field of face super-resolution. Future related researches should focus on multi-modal information decision fusion and tensor fusion to improve the feature dimension and temporal similarity of the reconstructed images;enhance the generalization ability of the model through large-scale pre-training and adversarial learning;investigate the impact of identity consistency algorithms and technologies such as transfer learning on complex imaging conditions.更多还原