| Résumé: | Low-light is an inescapable element in daily surroundings that greatly affects the efficiency
of human vision. However, current studies in low-light fundamentally lack an indepth
understanding of natural vision in low-light that would strengthen the development
of effective algorithms. This has subsequently restricted the development of well-rounded
systems that would aid in low-light environments, such as assistive systems, surveillance,
and autonomous car driving. Therefore, this thesis aims to study low-light image data to
gain a better understanding of their characteristics, and then based on this understanding,
investigate a computer vision solution that would pave the way for the advancement of
future assistive systems to operate in low-light conditions. An obvious challenge faced in
this study is the lack of a go-to database in this domain, hence led to the first contribution
that is a collection of 7,363 low-light images gathered from multiple sources, with
12 object classes annotation in order to facilitate the analysis for the purpose of applications.
From this dataset, it was found that low-light environments can be categorized
into 10 illumination types, each with different global and local characteristics that could
have different impact on a system. The second contribution is an in-depth analysis of the
collected data, specifically, by studying the global and local pixel intensities, followed by
the performance and visualizations of hand-crafted and learned features. It is found that
characteristics of the low-light pixel intensities provide a great challenge to algorithms.
The design of conventional hand-crafted features are greatly rooted to the behaviors of
bright environments, that they are unable to adequately address noise and lack of details
accompanying low-light images. Whereas, learned features revealed that the same object
yields amply different features in bright and low-light conditions, and irregular illumi nation greatly challenges the attention of the said features. These insights prompt the
third contribution, to propose a low-light contrast enhancement algorithm that is not only
able to improve the visibility but more importantly to reveal informative features to assist
high level applications. To this end, the Gaussian Process is studied as the contrast
enhancement approach to model the complexity of the local luminance variations, the primary
difficulty in low-light images. Experimental results show that the proposed method
outperforms the state-of-the-art in the common visual quality measure, the peak signalto-
noise ratio (PSNR) by 1.17dB. Additionally, novel information retrieval measurements
are proposed to better evaluate the usefulness of enhancement algorithms in applications,
namely the local features matching and l1-norm distance measure of intensity histogram.
Both of which the proposed method outperforms the state-of-the-art method by a large
margin, signifying the applicability of the proposal to support computer vision systems.
As a whole, the contributions of this study will push forward the advancement of computer
vision towards practicality in low-light environments which will be particularly
valuable in the development of assistive and surveillance systems that ensure the quality
of life and safety of the public
|
|---|
