Capsule GAN for robust face super resolution
- PDF / 1,434,234 Bytes
- 14 Pages / 439.642 x 666.49 pts Page_size
- 82 Downloads / 186 Views
Capsule GAN for robust face super resolution Mahdiyar Molahasani Majdabadi1 · Seok-Bum Ko1,2 Received: 26 March 2020 / Revised: 5 July 2020 / Accepted: 28 July 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Face hallucination is an emerging sub-field of Super-Resolution (SR) which aims to reconstruct the High-Resolution (HR) facial image given its Low-Resolution (LR) counterpart. The task becomes more challenging when the LR image is extremely small due to the image distortion in the super-resolved results. A variety of deep learning-based approaches has been introduced to address this issue by using attribute domain information. However, a more complex dataset or even further networks is required for training these models. In order to avoid these complexities and yet preserve the precision in reconstructed output, a robust Multi-Scale Gradient capsule GAN for face SR is proposed in this paper. A novel similarity metric called Feature SIMilarity (FSIM) is introduced as well. The proposed network surpassed state-of-the-art face SR systems in all metrics and demonstrates more robust performance while facing image transformations. Keywords Generative Adversarial Network (GAN) · Capsule network · Super resolution · Face hallucination
1 Introduction Face Super-Resolution(SR) is a fast-growing field that aims to enhance the resolution of facial images. These systems attempt to reconstruct High-Resolution (HR) face image from its Low-Resolution (LR) counterpart accurately. Due to the importance of facial details This work is the expansion of “MSG-CapsGAN: Multi-Scale gradient capsule GAN for face super-resolution,” in 2020 International Conference on Electronics, Information, and Communication (ICEIC), Barcelona, Spain, Jan. 2020. Seok-Bum Ko
[email protected] Mahdiyar Molahasani Majdabadi [email protected] 1
Department of Electrical and Computer Engineering, University of Saskatchewan, Saskatoon, Canada
2
Division of Biomedical Engineering, University of Saskatchewan, Saskatoon, Canada
Multimedia Tools and Applications
on human perception, it is vital to preserve these facial details [3]. Face hallucination has widespread and crucial applications in various face-related systems such as face recognition, video surveillance system, and image editing [11]. To reconstruct HR image accurately, several challenges should be overcome. First, for large scale face SR, reconstructing an accurate HR image is an arduous task due to the lack of information in the LR input. Second, it is required that the HR image not only possesses similarity to the ground-truth but also has a photo-realistic appearance and seems natural. Finally, faces can appear in unlimited different poses. Hence, the facial SR system should be pose-invariant to generalize for various situations. There are two categories of learning-based SR systems, local patch-based methods and global methods. In the first category, the system is trained to reconstruct a patch of an image at a time. Rajput and Arya propose mi
Data Loading...
