ProDeblurGAN: Progressive growing of GANs for blind motion deblurring in face recognition

Abstract

Elimination of motion blur is one of the key challenges in face recognition, especially from the video feed of low quality surveillance cameras. We present ProDeblurGAN, a progressively growing Generative Adversarial Nets (GAN) for single photograph motion deblurring. The current state-of-the-art DeblurGAN network performs well in structural similarity and visual appearance. However, the discriminator’s inability in distinguishing real and fake images concerning to finer or lower-level details result in the generator synthesizing low quality deblurred image. With recent advances in the progressive growing of GANs, we propose a motion deblurring model to utilize the advantages of such progressive adversarial training and generate a high quality deblurred image. The approach to gradually increase layers in training instead of initializing all at once facilitates the generation of higher resolution images. Also, at higher resolution, the finer details like the facial hair, moles, freckles and so on, are considered as features and used in the learning process. The proposed generator takes the blurred image as input as opposed to random noise in the case of traditional GAN networks. Implementation of the proposed method in face recognition systems will enhance the quality of the captured image and the accuracy of recognition of the face. The quality of the deblurred image generated by ProDeblurGAN is evaluated using state-of-the-art metrics.