Spatial and Transform Domains Content-Specific Image Quality Metrics with Applications for Biomedical and Security Imaging Systems
Samani, Arash.
2019
-
Information extraction
from digital images and videos are transforming and having a tremendous impact in every
discipline for their ability to help automate and make more accurate decisions in
systems applications. This includes improved detection and rapid diagnosis in biomedical
applications, more robust security and safety systems, improving automation in
manufacturing and industrial ... read moreproduction. Addressing challenging problems involving large
scale data including environmental monitoring and forecasting, and strategic farming
practices to improve lives and conserve natural resources. In any vision system
application, the images and videos often require evaluating the quality of an image and
its suitability for use in applications. Image quality measures can be used for object
detection or system quality evaluation for acceptance or rejection, for enhancement
algorithms, system validation, quality of service, and integrity verification. Humans
are very good at such subjective quality assessment of images, but as the number of
digital images increases, it is becoming practically impossible to use human assistance
for image quality evaluation for large scale data applications. This seeded the question
of whether it was possible to implement mathematical algorithms that can evaluate image
quality from the perception of human vision. Traditional subjective image metrics or
measures were developed to provide subjective quality assessment that can replace human
intervention in image evaluation. However, the expansion of artificial intelligence and
autonomous systems, and the increasing processing power of computers, are reducing human
interaction with everyday technology. This increases the demands for use of image
quality measures in machine vision and artificial intelligence applications. Originally,
the image measures were used only for image quality evaluation, but their use could be
expanded to optimize enhancement parameters. Based on these observations, we
hypothesized that the performance of an image quality measure is highly dependent on the
1) image type and 2) contents, 3) distortion type, and 4) the enhancement algorithm in
which it is used. The practice of choosing the correct measure for the class of test
images was previously accomplished by evaluating similar images with various measures
and compare the results to mean opinion score (MOS), which requires human evaluators.
However, using MOS becomes unrealistic as new data intensive applications become
prevalent including remote sensing, biomedical imaging sensors, unmanned aerial vehicles
(UAV), and satellite imagery. From our experiments, we observe that different image
distortions affect the image contents differently. Furthermore, because different
measures do not explore a mathematical relationship between image properties the same
way, their evaluation of the same image may vary for different distortions. These gave
motivation for this dissertation to introduce measures for specific classes of images
and distortions. This thesis provides alternative methods to find the appropriate
measure for different class of images without requiring MOS, thus saving massive human
labor to generate the MOS. With advances in machine learning and computer vision,
technology moves towards more autonomy and minimize human interaction in autonomous
systems. This thesis aims to contribute to the goal of reducing the need for human
intervention in image evaluation by introducing new measures and new methodologies for
measure selection. It also introduces applications for image quality metrics in image
enhancement and video analysis and new image enhancement algorithms. For example, we
introduce the first no-reference transform domain image quality metrics for gray scale
and color images. Other contributions of this thesis are (a) to use the image measures
for parameter optimization during the image enhancement, (b) introducing new image
enhancement algorithm in transform domain to increase the measure outcome, (c) using
image measures in video analysis for self-diagnostics of surveillance
systems.
Thesis (Ph.D.)--Tufts University, 2019.
Submitted to the Dept. of Electrical Engineering.
Advisor: Karen Panetta.
Committee: Karen Panetta, Sos Agaian, Aruna Ramesh, and Ronald Lasser.
Keyword: Engineering.read less - ID:
- x920g918g
- To Cite:
- TARC Citation Guide EndNote
