Convolutional neural network for inverse problems in image processing

In image processing, inverse problems pertain to the process of inferring an observed image to restore the version of the image before it experienced degradation. The objective of restoration is to remove or reduce a degradation that has occurred as well as to reconstruct or recover an image that has been corrupted by using a prior knowledge of the degradation phenomenon. Restoration techniques are oriented towards modelling the degradation and applying the inverse process in order to enhance or improve the quality of an image.
Regularized iterative algorithms have emerged as the standard approach to solving inverse problems in the past few decades. However, a more recent trend is the use of deep learning based algorithms, which has arisen as a promising framework providing the state-of-the-art performance for specially computer vision and image processing applications. The motivation of deep learning is that practically all aspects of the model are directly learnt from the data, starting with the lowest-level features presenting a suitable representation of the data and progressively providing higher-level abstractions for specific problems as one advances through the different network layers. Central to this technique has been the convolutional neural network (CNN) architectures.
The aim of this thesis to investigate the capability of CNN of resolving some inverse problems with a focus on low-level image processing tasks including super-resolution (SR), over-exposure correction, demosaicking and denoising. We propose a novel, versatile convolutional neural network framework, which can be effortlessly tailored to handle the specific task. We treat each task as a plain discriminative learning problem involving a feed-forward CNN, which directly outputs the latent clean image. Results of detailed experiments conducted and the comparison of performance with state-of-the-art approaches demonstrate that our framework is robust, effective and efficient in nature.
The key focus of the research conducted within this thesis is designing CNN architectures that are less complex, are designed tailored for a task in mind and will take less data and time for training, therefore making such architectures useful in device edge computing applications. We initially present a novel and promising approach to achieving single image super resolution (SISR) by proposing a novel self-feature, network loss function. The proposed networks performance was compared subjectively, objectively and computationally with the state-of-art in CNN based super-resolution image creation proving significant novel contributions to the subject area that is enabled by this work. Secondly, by the effective use of a synthesized training dataset, we propose a novel CNN framework that can be used to accurately correct over-exposure issues in digital photographs. The proposed networks performance was compared subjectively, objectively and computationally with the state-of-art in CNN based over-exposure correction approaches proving significant novel contributions to the subject area that is enabled by this work. Thirdly, the thesis proposes a novel CNN framework for image demosaicking, that makes use of the knowledge of Colour Filter Array used in image mosaicking to enable accurate image demosaicking but needing significantly less training data as compared to the state-of-art algorithms. The proposed networks performance was compared subjectively, objectively and computationally with the state-of-art in CNN based image democaicking proving significant novel contributions to the subject area that is enabled by this work. Finally, we propose a novel CNN architecture for blind noise removal, an approach much superior to existing CNN based noise removal architectures that are trained specifically to remove one type of noise. Results for images corrupted by different types of noise demonstrate that BiDNet is competent at preserving details of original images and outperforms other state-of-the-art methods in terms of qualitative and quantitative evaluations.