Deep Learning for Robust Super-resolution

Abstract

Super Resolution (SR) is a process in which a high-resolution counterpart of an image is reconstructed from its low-resolution sample. Generative Adversarial Networks (GAN), known for their ability of hyper-realistic image generation, demonstrate promising results in performing SR task. High-scale SR, where the super-resolved image is notably larger than low-resolution input, is a challenging but very beneficial task. By employing an SR model, the data can be compressed, more details can be extracted from cheap sensors and cameras, and the noise level will be reduced dramatically. As a result, the high-scale SR model can contribute significantly to face-related tasks, such as identification, face detection, and surveillance systems. Moreover, the resolution of medical scans will be notably increased. So more details can be detected and the early-stage diagnosis will be possible for many diseases such as cancer. Moreover, cheaper and more available scanning devices can be used for accurate abnormality detection. As a result, more lives can be saved because of the enhancement of the accuracy and the availability of scans.

In this thesis, the first multi-scale gradient capsule GAN for SR is proposed. First, this model is trained on CelebA dataset for face SR. The performance of the proposed model is compared with state-of-the-art works and its supremacy in all similarity metrics is demonstrated. A new perceptual similarity index is introduced as well and the proposed architecture outperforms related works in this metric with a notable margin. A robustness test is conducted and the drop in similarity metrics is investigated. As a result, the proposed SR model is not only more accurate but also more robust than the state-of-the-art works.

Since the proposed model is considered as a general SR system, it is also employed for prostate MRI SR. Prostate cancer is a very common disease among adult men. One in seven Canadian men is diagnosed with this cancer in their lifetime. SR can facilitate early diagnosis and potentially save many lives. The proposed model is trained on the Prostate-Diagnosis and PROSTATEx datasets. The proposed model outperformed SRGAN, the state-of-the-art prostate SR model. A new task-specific similarity assessment is introduced as well. A classifier is trained for severe cancer detection and the drop in the accuracy of this model when dealing with super-resolved images is used for evaluating the ability of medical detail reconstruction of the SR models. This proposed SR model is a step towards an efficient and accurate general SR platform.