Skip to Main Content
Frequently Asked Questions
Submit an ETD
Global Search Box
Need Help?
Keyword Search
Participating Institutions
Advanced Search
School Logo
Files
File List
43007.pdf (6.61 MB)
ETD Abstract Container
Abstract Header
Quality Aware Video Processing for Deep Learning Based Analytics Tasks
Author Info
Ikusan, Ademola
Permalink:
http://rave.ohiolink.edu/etdc/view?acc_num=ucin1659517930534801
Abstract Details
Year and Degree
2022, PhD, University of Cincinnati, Engineering and Applied Science: Computer Science and Engineering.
Abstract
Wireless video systems have been heavily deployed for various distributed sensing purposes, such as attempting to detect, recognize, track objects, and understand their behaviors which have been integrated into surveillance to provide real-time analysis results to human operators who will make final decisions. As technology advances, we have seen deep neural networks be a backbone technology supporting modern intelligent mobile applications and they have the ability to perform highly accurate and reliable inference tasks. Wireless cameras have played a big role in video surveillance capacity and are equipped with an embedded camera with video capture, video encoding or local video processing, and data transmission. The process of video analysis is implemented either in the central server or in the sensor node, depending on their computational capability, energy supply, and the purpose of applications. We have studied two paradigms used in video analytics, the first one is extracting the visual information at a wireless camera node and then the information is sent to a central processing hub for processing, this paradigm is commonly known as the extract-compress-analyze paradigm which is the traditional strategy. While the second paradigm extracts visual information using a wireless camera and it partially analyzes the visual information before compressing and sending it over to the processing hub for final computation and use cases, this paradigm is commonly known as the extract-analyze-compress strategy which is a recent strategy. Firstly, we studied on the extract-compress-analyze strategy. We investigated the impact of distortions such as noise, blur, etc. on visual information generated from wireless cameras. Based on studies, distortions such as noise, blur, bad lighting, etc. are introduced to visual information at the point of generation which negatively impacts computer vision applications such as object detection, classification, etc. Based on these observations, we proposed a quality assessment and adjustment framework for automatic video analytics systems to improve the quality of the performance of computer vision applications. Furthermore, based on the second paradigm which is the extract-analyze-compress paradigm, we investigated the impact of lossy compression on deep features. It has been found that deep features of deep neural network algorithms are large and can be computationally intensive to produce. In terms of computation, processor chips have been specifically made for portable devices to handle such computation, especially partially. An aspect that needs more research is how the feature maps can be compressed and the impact of the compression on computer vision tasks. This finding brought about three objectives: one is to create a compression framework that can make use of the redundant information present in the feature maps for compression and evaluate it on a computer vision task like image classification; the second is to develop a distortion model for distortion introduced by the double compression used in the framework and; lastly to see the trade-off between the quality of computer vision tasks and the encoding rate optimized.
Committee
Rui Dai, Ph.D. (Committee Member)
Gowtham Atluri, Ph.D. (Committee Member)
Heng Wei, Ph.D. (Committee Member)
Boyang Wang (Committee Member)
Anca Ralescu, Ph.D. (Committee Member)
Pages
102 p.
Subject Headings
Computer Science
Keywords
Computer Vision
;
Compression
;
Rate-Distortion Optimization
;
Deep Learning
;
Machine Learning
;
Image Quality Assessment
Recommended Citations
Refworks
EndNote
RIS
Mendeley
Citations
Ikusan, A. (2022).
Quality Aware Video Processing for Deep Learning Based Analytics Tasks
[Doctoral dissertation, University of Cincinnati]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=ucin1659517930534801
APA Style (7th edition)
Ikusan, Ademola.
Quality Aware Video Processing for Deep Learning Based Analytics Tasks.
2022. University of Cincinnati, Doctoral dissertation.
OhioLINK Electronic Theses and Dissertations Center
, http://rave.ohiolink.edu/etdc/view?acc_num=ucin1659517930534801.
MLA Style (8th edition)
Ikusan, Ademola. "Quality Aware Video Processing for Deep Learning Based Analytics Tasks." Doctoral dissertation, University of Cincinnati, 2022. http://rave.ohiolink.edu/etdc/view?acc_num=ucin1659517930534801
Chicago Manual of Style (17th edition)
Abstract Footer
Document number:
ucin1659517930534801
Download Count:
40
Copyright Info
© 2022, some rights reserved.
Quality Aware Video Processing for Deep Learning Based Analytics Tasks by Ademola Ikusan is licensed under a Creative Commons Attribution-NonCommercial-NoDerivs 3.0 Unported License. Based on a work at etd.ohiolink.edu.
This open access ETD is published by University of Cincinnati and OhioLINK.