Implicit Neural Representations For Non-blind Depth-aware Image Deblurring

Abstract

The purpose of this thesis is to evaluate different implicit neural representation architectures in a self-supervised framework for image deblurring, where depth information is available. Image deblurring is the task of recovering sharp details from a blurry image. Classical methods for image deblurring often rely on computationally-expensive iterative algorithms that perform poorly on natural images, while fully-supervised deep learning approaches require large-scale datasets for training. Implicit neural representations have emerged as a powerful tool to represent multidimensional signals through neural networks. By coupling an implicit neural representation with a differentiable blur formation model, deep-learning optimization methods can be used to learn a neural representation that produces a sharp image, using only the blurry input for supervision. In particular, this thesis explores two computationally efficient hybrid implicit neural representations: 1) Instant Neural Graphic Primitives, and 2) Dictionary fields.

The thesis is divided into 3 parts, First going through the background theory behind image blur, image deblurring and implicit neural representations. Secondly an introduction of the hybrid implicit neural representations used for image deblurring in this thesis. Lastly the results obtained from the experiments will be analyzed and discussed.

The research done shows impressive results and promising possibilities for further optimization. The results gained from the state-of-the-art methods used have shown to be computationally more efficient and have been able to produce superior quality to the baseline models used as comparisons, outperforming them both in training time and result quality