Abstract |
The rapid evolution of display technologies and imaging sensors has raised an enormous
interest in the imaging and multimedia communities, and thus the demand for providing highquality images has grown tremendously. Computational photography and image processing
aim to improve the interpretability and perception of imaging content in order to facilitate
subsequent tasks, such as automated analysis, detection, segmentation and recognition,
among others. Despite the significant increase in the spatial resolution and the introduction
of 3D content, spectral resolution, dynamic range enhancement, image denoising and
deconvolution remain attractive research problems, as there is still room for improvement.
The key contribution of this thesis is the design of novel post-acquisition, machine learning
formulations that tackle the main limitations of imaging sensors, with special emphasis on
satellite imaging technologies and multi-and hyperspectral imaging systems.
In this dissertation, we propose novel and robust mathematical Coupled Dictionary Learning
formulations, based on the intelligent scheme of the Alternating Direction Method of
Multipliers (ADMM). We aim to learn dictionaries for sparse modeling in joint feature spaces
while enforcing the desired relationships among the different signal representations. The
proposed algorithmic formulations are able to overcome the main limitations of traditional
Coupled Dictionary Learning schemes. Thus we are addressing a twofold problem. First, we
extend the existing methodology of Sparse Representations and Coupled Dictionary Learning
which is very important from an engineering point of view. On the other hand, we are assisting
the research community which has started to appreciate the necessity of coupled feature
spaces in several signal and image processing tasks, from image super-resolution,
deconvolution and dynamic range synthesis of modern sensors, to hyperspectral and satellite
image enhancement to further investigate this research area with novel algorithmic
formulations.
The structure of this dissertation goes as follows. Primarily, we address the problem of
dynamic range enhancement of standard 8bit imagery. High Dynamic Range (HDR) imaging
technology is acknowledged as the next success story in consumer imaging, with a vastly
increasing image and video acquisition, content reproduction, and graphics applications
supporting HDR features. We produce HDR content from a single image using an innovative
Machine Learning (ML) approach that first generates a sequence of bracketed exposures via
a Joint Dictionary Learning formulation, encoding the transformations between differently
exposed images, and then combines them ideally into a well-illuminated low dynamic range
(LDR) image or merges them directly into a HDR image. Additionally, we learn appropriate
features by employing a stacked sparse autoencoder (SSAE) based framework.
Second, we confront one of the major limitations of multi- and hyperspectral imagery: the
spectral resolution enhancement. The majority of literature approaches mostly focuses on the
enhancement of spatial resolution, and thus only a handful of techniques tackle the problems
of spectral and spatio-spectral resolution enhancement of satellite sensors. In this
dissertation, we develop a coupled dictionary learning model which considers joint feature
spaces, composed of low and high spectral resolution hyper-cubes, in order to achieve
spectral super-resolution performance. We formulate our spectral coupled dictionary learning
optimization problem within the context of the Alternating Direction Method of Multipliers,
and we manage to update the involved quantities via closed-form expressions.
Moreover, we investigate the core problem of spectroscopic data denoising on simulated
Euclid-like noisy templates, and we exploit the proposed ADMM Coupled Dictionary Learning
methodology in order to learn coupled feature spaces, composed of high- and low-quality
spectral profiles. The reconstructed spectral profiles are able to improve the accuracy,
reliability and robustness of automated redshift estimation methods.
Additionally, another significant contribution of this dissertation includes the deconvolution
of astronomical imagery. Specifically, we design a novel and robust post-acquisition
formulation that recovers the high-quality versions of blurry astronomical observations, along
with information details regarding the blur kernel, i.e. point spread function (PSF). In order to
accomplish this goal, we exploit the mathematical frameworks of Sparse Representations, and
the Alternating Direction Method of Multipliers.
Likewise, we apply our proposed ADMM Coupled Dictionary Learning scheme on a highly
challenging problem of the remote sensing community. Specifically, we consider the retrieval
of the active measurements of the Soil Moisture Active Passive (SMAP) satellite. The upper
goal of this study is the direct soil moisture estimation from SMAP's satellite measurements.
In the third part of this dissertation, we consider a novel hyperspectral image understanding
technique. Specifically, we propose a novel machine learning technique that addresses the
hyperspectral image classification problem by employing the state-of-the-art scheme of
Convolutional Neural Networks (CNNs). The formal approach introduced in this work exploits
the fact that the spatio-spectral information of an input scene can be encoded via CNNs and
combined with multi-class classifiers. The proposed deep feature learning scheme is focused
on the classification of snapshot mosaic hyperspectral imagery, while a new hyperspectral
classification dataset of indoor scenes is constructed.
Last but not least, we have conducted a complete investigation of the experimental setups
with multiple real-world's and synthetic datasets and compare it with the best state-of-theart algorithms.
|