基于相似性约束的面部幻象在人脸识别中的应用

为了根据低分辨率（LR）人脸图像生成高分辨率（HR）图像以提高人脸识别率，设计了四个相似性约束函数，提出了基于相似性约束的面部幻象方法。首先利用LR-LR约束计算出输入的LR人脸图像与训练集中各LR人脸图像之间的相似性；然后利用LR-HR约束描述输入的LR人脸图像与HR训练图像之间的局部结构相似性，同时增强相邻幻象图像块之间的平滑约束；最后利用空间相似性约束减少远离幻象图像块的那些图像块的影响。在FERET、Yale及ORL三大通用人脸数据库上的实验结果表明，相比其它几种较为先进的面部幻想生成方法，所提方法得到的幻想图像分辨率更高，此外，所提方法得到了更高的识别率。

Face Hallucination Based on Similarity Constraints for Face Recognition

To generate high resolution image by low resolution face image so as to improving recognition accuracy, four similarity constraint functions are designed, based which face hallucination based on similarity constraints is proposed. Firstly, LR-LR constraint is used to compute similarity between inputted LR face image and that in training sets. Then, LR-HR constraint is used to describe local structure similarity between LR face image and HR training images, and slider constraint within neighboring hallucination image blocks. Finally, the spatial similarity constraint is used to reduce effect of image block away from those image blocks. Experimental results on FERET, Yale and ORL show that proposed method has hallucination image with higher resolution and get better face recognition accuracy than several advanced face hallucination approaches.