Date of Award

5-2018

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

Committee Chair

Jan P. Allebach

Committee Co-Chair

Qian Lin

Committee Member 1

Amy R. Reibman

Committee Member 2

Fengqing M. Zhu

Committee Member 3

George T. Chiu

Abstract

In this dissertation, studies on digital halftoning and human facial landmark localization will be discussed. 3D printing is becoming increasingly popular around the world today. By utilizing 3D printing technology, customized products can be manufactured much more quickly and efficiently with much less cost. However, 3D printing still suffers from low-quality surface reproduction compared with 2D printing. One approach to improve it is to develop an advanced halftoning algorithm for 3D printing. In this presentation, we will describe a novel method to 3D halftoning that can cooperate with 3D printing technology in order to generate a high-quality surface reproduction. In the second part of this report, a new method named direct element swap to create a threshold matrix for halftoning is proposed. This method directly swaps the elements in a threshold matrix to find the best element arrangement by minimizing a designated perceived error metric. Through experimental results, the new method yields halftone quality that is competitive with the conventional level-by-level matrix design method. Besides, by using direct element swap method, for the first time, threshold matrix can be designed through being trained with real images. In the second part of the dissertation, a novel facial landmark detection system is presented. Facial landmark detection plays a critical role in many face analysis tasks. However, it still remains a very challenging problem. The challenges come from the large variations of face appearance caused by different illuminations, different facial expressions, different yaw, pitch and roll angles of heads and different image qualities. To tackle this problem, a novel coarse-to-fine cascaded convolutional neural network system for robust facial landmark detection of faces in the wild is presented. The experiment result shows our method outperforms other state-of-the-art methods on public test datasets. Besides, a frontal and profile landmark localization system is proposed and designed. By using a frontal/profile face classifier, either frontal landmark configuration or profile landmark configuration is employed in the facial landmark prediction based on the input face yaw angle.

Download

Share

COinS