基于图像风格迁移的人脸识别域适应方法

监控场景下的带标签人脸数据难以获取，尽管可以利用已有的公开数据集或合成数据，但这些数据与真实的监控人脸数据在图像风格上存在较大的域间差异。针对该问题，不同于基于特征或公共子空间的域适应方法，提出一种基于图像风格迁移的解决方法。具体地，基于CycleGAN网络改进得到Face-CycleGAN，在保持身份属性的前提下，对现有带标签数据进行风格迁移，使其在背景、光照、皮肤材质等方面与监控场景更接近，并进一步通过联合滤波对迁移图像进行后处理。最后，利用迁移得到的数据优化人脸识别算法，减小域间差异带来的负面影响。提出的方法在公开数据集EK-LFH和自建数据集3DProj-Sur上进行了实验评估，分别取得了21.93%和4.77%的识别率提升，证明了该方法在解决域适应问题上是有效的。

Domain adaption for face recognition based on image style transfer

It''s difficult to obtain labeled face data in video surveillance, although there are large public datasets or synthetic data, they differ significantly from surveillance face data on image style. To solve this problem, in contrast to existing domain adaptation methods based on features or public subspace, this paper proposed a new method based on image style transfer. Specifically, the method constructed Face-CycleGAN based on CycleGAN network, on the premise of maintaining identity consistency, it could transfer the style of labeled data to make it more similar to the data in video surveillance in terms of background, illumination, skin texture, etc. Combined filter further processed transferred images as a post-process step. Finally, the method trained a face recognizer by using the transferred results to narrow the gap between the domains. Experimental results on a public dataset EK-LFH and a self-collected dataset 3DProj-Sur show that the proposed method achieves 21.93% and 4.77% performance improvement respectively comparing to original model, which proves the effectiveness of the proposed method on domain adaption.