Self-Supervised Effective Resolution Estimation with Adversarial Augmentations

High-resolution, high-quality images of human faces are desired as training data and output for many modern applications, such as avatar generation, face super-resolution, and face swapping. The terms high-resolution and high-quality are often used interchangeably; however, the two concepts are not equivalent, and high-resolution does not always imply high-quality. To address this, we motivate and precisely define the concept of effective resolution in this paper. We thereby draw connections to signal and information theory and show why baselines based on frequency analysis or compression fail. Instead, we propose a novel self-supervised learning scheme to train a neural network for effective resolution estimation without human-labeled data. It leverages adversarial augmentations to bridge the domain gap between synthetic and real, authentic degradations - thus allowing us to train on domains, such as hu-man faces, for which no or only few human labels exist. Finally, we demonstrate that our method outperforms state-of-the-art image quality assessment methods in estimating the sharpness of real and generated human faces, despite using only unlabeled data during training.

1. Introduction

Generative models that produce high-quality face images, such as human avatar generation, face super-resolution, and face swapping, require high-resolution, high-quality training data sets. Low-resolution images neg-atively affect the resulting models' output. It is therefore crucial to identify and remove these images prior to training. This seemingly simple task is challenging, however, since high-resolution does not always imply high-quality, so users must often resort to visual inspection.