Learning-based image reconstruction for computational photography

Abstract

The emergence of digital imaging made the photography available for everyone by lowering the cost of imaging devices and eliminating the costs of films and the development processes for them. Along with it, a novel imaging application field, so-called computational photography, was introduced by combining digital imaging and computing. Since it enables to enhance the quality of images and to fix flaws in photographs afterward shooting, computational photography has been playing an essential role in digital imaging.

At the core of computational photography, there is a technical challenge to solve a severely ill- posed optimization problem of inverse light transport. Computational photography algorithms need to overcome the mathematical challenge to achieve high-quality image output. With efforts of addressing the ill-posedness of the inverse optimization problem, many hand-crafted image priors have been commonly used to regulate the range of possible solutions, yielding plausible high-quality images. However, traditional approaches cannot handle the broad spectrum and diversity of real-world images.

In this thesis, I demonstrate three learning-based image reconstruction methods for computational photography, based on priors obtained by learning from image data: compressive hyperspectral imaging, high dynamic range (HDR) imaging, and novel view synthesis. First, for compressive hyperspectral imaging, I introduce a novel spectral prior, which is learned using a deep convolutional autoencoder. Second, for high dynamic range imaging, I adopt joint sparse coding to acquire a prior knowledge on HDR images. Finally, for the novel view synthesis method, I propose a new approach that combines a conventional view synthesis and a learning-based refinement using a convolutional neural network. Qualitative and quantitative comparisons demonstrate that the learned priors are highly effective in addressing each of the three problems, validating that the three proposed methods outperform state-of- the-art techniques with high image quality.