GAN-Based Two Stage Network for De-occlusion Face Image

In deep learning applications, facial recognition systems such as facial recognition, facial emotion recognition, facial-based criminal investigation, and facial-based driver monitoring have been researched. These applications show very high performance in recognition and analysis, but there is a problem that the performance is degraded when the face is covered. The purpose of this paper is to restore facial occlusion covered by obstruction. An image generation network is used to restore the face from object such as glasses, makeup, masks, and intentional obstruction. The image generation network recognizes facial occlusion in the input image and fills the area with as natural a look as possible. The proposed model is a three-stage model, and in the first step, the occluded part is recognized and then the occluded object is separated. In the second step, two generators are connected to generate an image from which the occluded object is removed. We used two discriminators to determine the overall appearance and the hidden part, and the resulting results affect the generator’s learning compared to the correct answer image. The model’s learning uses the original image and the occulded image dataset using CelebA and FFHQ. For the performance analysis of the model, we compared the performance of the proposed model with the comparative models using Fréchet inception distance (FID), SSIM and Peak Signal to Noise ratio (PSNR) and the proposed model show 0.8 lower at FID and 0.012 higher at SSIM compared to other models.