Computational Image Aesthetic: Measurement, Analysis and Applications
BAO, LONG.
2019
-
Existing image-related
research has advanced on many levels including the signal level, such as image
denoising, image enhancement, and the content level including image segmentation, image
classification, and image summation. However, limited research has focused on
aesthetic-level image analysis of a human's aesthetic perception over digital
photography. To meet this gap, this dissertation ... read moreaims at addressing three specific image
aesthetic-related questions: 1) What kind of element or pattern within an image makes it
pleasing to a human?; 2) How do we develop algorithms to do enhancement of an image at
the aesthetic level?; and 3) How can we develop algorithms to enable computer vision to
evaluate the quality of an image?. Based on the study of different image
aesthetic-related elements, this dissertation focuses on three important elements: the
color element and the illuminance element, which have positive influences on enhancing
the aesthetic level of an image, and the distortion element, which shows the negative
influence on destroying an image's aesthetic atmosphere. Considering the color element,
two new types of color transfer algorithms, including global color transfer and
selective color transfer, and a new color theme-based aesthetic enhancement algorithm
are proposed to achieve flexible, effective and fast image aesthetic enhancement and
processing. Considering the illuminance element, this paper introduces a new
contrast-based aesthetic enhancement approach based on adding the sense of the depth to
the scene within an image. Considering the distortion element, a new concept of
sequence-to-sequence similarity is introduced for noise distortion removal, and
corresponding noise removal algorithms are developed to generate high visual-quality
images as the prerequisite of an aesthetic image. To quantitatively, objectively and
automatically measure the quality of an image, this dissertation analyzes the human's
behavior over image quality measure/comparison, and, thus, introduces a new unsupervised
image quality metric and a new supervised neural network-based image quality comparator.
All these developed algorithms in the research of image aesthetic show significant
potential in many practical applications, including fashion design, support of
color-blind people, camouflage pattern design support, camouflage detection system,
computational photography, medical imaging systems, and computer
animation.
Thesis (Ph.D.)--Tufts University, 2019.
Submitted to the Dept. of Electrical Engineering.
Advisor: Karen Panetta.
Committee: Sos Agaian, Liping Liu, and Brian Tracey.
Keyword: Electrical engineering.read less - ID:
- 2227n251j
- To Cite:
- TARC Citation Guide EndNote
