Image Restoration Using Deep Learning

Abstract

In this thesis, we propose several convolutional neural network (CNN) architectures with fewer parameters compared to state-of-the-art deep structures to restore original images from degraded versions. Employing fewer parameters corresponds to a new trend in deep learning that seeks to create lighter/smaller models without affecting performance (i.e., quality of outcomes). Our models are used for image restoration tasks, such as single image super-resolution (SISR), denoising and artefacts reduction. Image restoration is a fundamental application in computer vision which is used in several practical fields such asmedical imaging and security systems. Recently, several state-of-the-art algorithms have been developed to infer high-resolution (HR) images from only low-resolution (LR) images using deep learning algorithms. Tackling distortion such as blurring in addition to downsampling in LR images is significant, although it has received much less attention. In this work, we proposemultiple deep learning architectures for simultaneously deblurring and producing HR images fromblurred and down-sampled images, which is considered a more challenging problem in image super-resolution. Two situations are considered: non-blind, where the nature and level of noise are known; and blind, where less information about the blurring process is available. The non-blind method uses a specific blur level while training the model and testing data, and this method is used when we have prior knowledge of the blur kernel. Also, it requires training a separate model for each blur kernel. However, when the blur kernel is not known, when the blur level is different for each image, the blind approach is used instead, where a single deep learning model is trained to restore a LR and blurry image. Furthermore, image compression formats such as MPEG and JPEG can present a combination of degradation effects in images such as blurring, ringing and blocking artefacts. We propose an approach that mitigates this undesirable compression drawback based on the use of CNN models. Our solution improves the visual quality of degraded images by automating the correct of any such artefacts. We also experimentally show that our proposed CNN architectures can give the same or slightly better results compared to the existing deeper structures that aremore costly computationally. Finally, we employ one of the proposed CNNmodels (DBSR) for denoising the natural noise due to low light conditions in a RENOIR dataset [12], to assess our model using real noisy images.