| dc.description.abstract | Image super resolution is one of the most significant computer vision researches
aiming to reconstruct high resolution images with realistic details from low resolution
images. In the past years, a number of traditional methods intended to produce
high resolution images. Recently, Deep Convolutional Neural Networks (DCNNs)
have developed rapidly and achieved impressive progress in the computer vision area.
Benefiting from DCNNs, the performance of image super resolution has improved
compared with traditional methods. However, there still exists a large gap between
the results of current methods and the real-world high resolution quality.
In this thesis, we leverage the techniques of DCNNs to develop image super res-
olution models for generating satisfactory high resolution images. There are several
proposed methods in this thesis to satisfy different super resolution scenarios. Our
proposed methods are based on Generative Adversarial Networks (GANs), leading
to powerful generative ability and effective discriminative learning. To breakthrough
current bottlenecks, we design novel architectures for generator and discriminator,
and involve new optimization strategies to improve the learning stability of the mod-
els. In order to improve the generalization ability of proposed methods, we conduct
two mainstream super resolution tasks, namely face image hallucination and natu-
ral image super resolution. All the proposed components of our methods result in
promising super resolution performance for these tasks.
Not only handling the supervised super resolution task, we also investigate the
more challenging problem, namely the unsupervised image super resolution task where
the paired high resolution image and low resolution image data are unavailable. To
evaluate the performance of our methods in different scenarios, we conduct exten-
sive experiments on several benchmark datasets to study each method separately.
Compared to state-of-the-art methods, our methods are able to achieve superior per-
formance both quantitatively and qualitatively. | en_US |
