Recently, significant advances in the field of automated face recognition have
been achieved using computer vision, machine learning, and deep learning
methodologies. However, despite claims of super-human performance of face
recognition algorithms on select key benchmark tasks, there remain several
open problems that preclude the general replacement of human face recognition
work with automated systems.
State-of-the-art automated face recognition systems based on deep learning
methods are able to achieve high accuracy when the face images they
are tasked with recognizing subjects from are of sufficiently high quality.
However, low image resolution remains one of the principal obstacles to face
recognition systems, and their performance in the low-resolution regime is decidedly
below human capabilities. In this PhD thesis, we present a systematic
study of modern automated face recognition systems in the presence of image
degradation in various forms. Based on our findings, we then propose a novel
technique for improving the quality of low-resolution face images. Specifically, we present a novel deep learning model architecture for image superresolution,
and a novel training procedure for face hallucination that trains
the model to super-resolve face images in a manner that preserves the information
about the subject identity present in the low-resolution image. We
validate the model by comparing its image reconstruction capability against
several state-of-the-art models, as well as its performance on downstream
semantic tasks including face recognition and face landmark localization.
Next, we study the generalization capabilities of super-resolution-based
face hallucination models, and find most of the models studied to be heavily
biased towards the articial image degradation process used to generate their
training datasets. We notice that due to this bias, none of the face hallucination
models considered are able to outperform an interpolation baseline
on face recognition benchmarks with real-life low resolution images.
To overcome this problem, we then develop a novel method for face recognition
from low-resolution images that uses the results of multi-scale face
hallucination models developed earlier. The proposed method is able to
benefit from the high-resolution information added by the face hallucination
models without suffering from the training set bias they exhibit, and systematically
outperform the interpolation baseline and other state-of-the-art
low-resolution face recognition models on the SCFace benchmark.
Our proposed methods are trained on large face image datasets in a manner
typical for deep learning models. However, the resulting trained models
are useful for face recognition applications in an open-set regime, and do not
need to be re-trained for novel subjects.
|
