The science of super-resolution imaging involves the construction of a single highresolution
representation by registering and fusing together multiple low-resolution
representations.These candidate representations may include images and videos.
This thesis covers four studies that advance the state-of-the-art in super-resolution
imaging.
The first study is based on the innovative idea of computing Event Models to
represent activities in the low resolution and low frame-rate input sequences. These
Event Models are then used, instead of the sequences, for temporal registration. Experimental
results show that the use of Event Models produces significantly better
reconstruction than contemporary approaches.
Current techniques in super resolution imaging limit the input video sequences
to be from the same scene or the same instance of an event. We propose a novel
method for generation of high-resolution video from sequences acquired from repetitions
of the same activity. The input video sequences may have different temporal
scales and spatial viewpoints. Our proposed method uses Event Model techniques
from the previous study and computes sequence synchronization to sub-frame accuracy.
We also demonstrate the application of this new method by constructing a
single 4D MRI sequence using multiple low resolution sequences.
MRI acquisition involves a fundamental trade-off between image quality and
frame rate. In our review, we found that current MRI technology can only capture
about seven frames per second before the image quality degrades below an acceptable
threshold. The low frame rate limits the usefulness of the MRI in the study
of high speed events, such as swallowing reflexes and cardiac motion. MRI can
be used to capture multiple instances of the same event from multiple view points.
These sequences are temporally offset, non-uniformly scaled, from different view
points and represent repetitions of the same event. Our third study builds on the
techniques developed in the previous studies to register these orthogonal MRI sequences.
This provides an ability to simultaneously view the orthogonal image
planes from a single arbitrary view point.
In our final study we address the problem of identifying a subset of low resolution
sequences, from a large set of samples, such that the resulting high resolution
video has the best possible reconstruction accuracy. To compare multiple input
video sequences, we first calculate an a priori confidence measure for each pair of
video sequences. The confidence measure is derived from the registration error and
non-uniformity of sequence samples. This confidence measure is a measure of diversity
in the selected subset of low resolution sequences. We then use our iterative
ranking algorithm to rank the low resolution sequences so that a minimum number
of sequences result in desired reconstruction accuracy.
Experimental results using
a wide variety of input video and audio sequences show that the reconstruction
accuracy of the proposed method is better than the reconstruction from a random
selection of input sequences.