Remove The Glasses From The Face Image! EyeGlassesGAN, Which Prevents Reduction In Face Recognition Performance By Eliminating A Shield From The Face

3 main points

✔️ Proposal for a framework that automatically removes glasses from facial images
✔️ Proposed data synthesis method for learning framework, tinted lenses/glare, etc.
✔️ Applying this framework to the preprocessing of face recognition significantly improves the accuracy of face recognition when wearing glasses

ByeGlassesGAN: Identity Preserving Eyeglasses Removal for Face Images
written by Yu-Hui Lee, Shang-Hong Lai
(Submitted on 25 Aug 2020)
Comments: Accepted at ECCV2020
Subjects: Multimedia (cs.MM)
 

overview

Facial recognition services are used in a variety of applications, including airport and online identification (eKYC) and smartphone unlocking (Face ID). State-of-the-art face recognition technology based on deep learning has made it possible to put these services to work. However, even today, facial recognition accuracy is reduced when the glasses cover part of the face.

In order to eliminate the influence of glasses on face recognition, we have been researching technology to remove them and Research has been conducted to "improve the recognition accuracy for glasses by improving the lack of training data when wearing glasses". Research on eyeglass removal techniques has been conducted using PCA and the subspace method, but the performance was insufficient and the validation in the face recognition task was inadequate. In addition, In a study on improving the lack of training data, a method for generating face images of people wearing glasses using GANs was studied and reported to be highly accurate. 

In this paper, we propose a new framework that removes the glasses while using GANs, which have been reported to be highly accurate. By using this framework for preprocessing face recognition, we found that the recognition accuracy was significantly improved when the glasses were worn. In addition, the proposed framework differs from the conventional GAN method in that it can generate facial images without glasses, so it can be used for various applications, such as virtual makeup for trial purposes.

Takumu