Date of Award

8-29-2019

Document Type

Dissertation - SCU Access Only

Publisher

Santa Clara : Santa Clara University, 2019

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering

First Advisor

Tokunbo Ogunfunmi

Abstract

The pervasiveness of hand-held battery-operated devices with various types of high quality video streams to and from them create the ever increasing need for high quality and low power video compression, which can also ensure low bandwidth, less storage translating also into lower cost. The applications vary from social media, to entertainment, education, health care and science, The state of the art video compression standard High Efficiency Video Coding (HEVC), developed by both the Motion Picture Expert Group (MPEG) and the Video Coding Expert Group (VCEG) reduces the required storage and bandwidth by up to 50% compared to the previous most widely deployed video coding standard H.264/AVC. However, it has significantly higher complexity than the previous H.264 coding standard and thus results in greater power consumption. The major contributor to the complexity (~70%) and to the power consumption is motion estimation, which renders it to be the main target for any power and efficiency enhancements. All new algorithms should be preferably preserving backward compatibility with the previous video coding standards.

In this dissertation we propose methods, which provide the application with full control over the quality-power trade-off.

First, we propose a signature-based algorithm for Fast Motion Estimation (FME), which targets Full Search (FS) quality at the speed of FME and also in real time.

Second, we develop enhanced fixed Search Pattern (SP) FME Algorithms, which are more adequate to HD video sequences (examples: Binary, Fibonacci and a Custom one).

Third and the most important proposed method in terms of novelty and impact to the field is the Adaptive method, which extracts multiple SPs from actual sample video sequences and builds a dictionary of such SPs then uses AI and its generalization properties to select the best SP to be used for FME for a given region of a video frame FME. Two more optimized algorithms have been also proposed for more cost effective implementations. The first one keeps the SP Dictionary, but replaces the AI function with existing logic and the second one removes even the dictionary storage, while preserving the adaptive nature of the algorithm.

The results of our simulations show close to Full Search quality (typically within 1dB) for the main Adaptive algorithm (with a dictionary and with AI [NN]).

Some of the other main contributions of this dissertation are the four rare properties of the proposed adaptive algorithms. They are first agnostic to resolution, second agnostic to frame rate, then they are equally applicable to both HEVC and H.264 video encoding standards and finally the search points in the adaptive search patterns are sorted/ordered according to their expected quality contribution, so in case of an early search termination (with the help of a certain threshold), the calculations, which have already happened are expected to have contributed the most to the total FME quality.

Download

Share

COinS