자료유형  

 학위논문

Control Number  

0016935424

International Standard Book Number  

9798380729819

Dewey Decimal Classification Number  

758

Main Entry-Personal Name  

Zhang, Zhuomin.

Publication, Distribution, etc. (Imprint  

[S.l.] : The Pennsylvania State University., 2023

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2023

Physical Description  

1 online resource(143 p.)

General Note  

Source: Dissertations Abstracts International, Volume: 85-05, Section: A.

General Note  

Advisor: Wang, James Z.;Li, Jia.

Dissertation Note  

Thesis (Ph.D.)--The Pennsylvania State University, 2023.

Restrictions on Access Note  

This item must not be sold to any third party vendors.

Summary, Etc.  

요약The rapid development of imaging technologies has revolutionized image acquisition and led to an unprecedented explosion in the volume of visual data, which has become an integral part of our daily lives. Furthermore, in most real-world scenarios, data is often collected from different sources, referred to as different "domains," each with its own characteristics and properties. Recently, the utilization of advanced machine learning (ML) algorithms to analyze cross-domain images has gained significant attention in various research fields. Cross-domain image analysis aims to: (1) identify and characterize similarities and discrepancies among images belonging to different domains; and (2) transfer knowledge learned from one domain to another domain for further image analysis tasks, which can help to resolve the issues of limited labeled data and domain-specific variations. Over the past decades, researchers have developed various techniques and methods to attain the objectives of cross-domain image analysis, which include domain adaptation, transfer learning, and deep-learning-based methods. Motivated by the superiority of these devised analysis approaches, this dissertation aims to advance the state-of-the-art in cross-domain image analysis, contributing to the development of more effective and robust machine learning models that can benefit a wide range of research fields.In this dissertation, we present two research problems and corresponding solutions from the biomedical and art historical perspectives to demonstrate the capability of advanced machine learning algorithms and underlying challenges when dealing with cross-domain images analysis. The first part discusses the utilization of the idea of image translation to achieve placenta segmentation from photos collected from different sources. For each photo, four regions need to be segmented, placenta disc, umbilical cord, ruler, and background. Specifically, we propose a multi-region saliency-aware learning (MSL) method for cross-domain placenta image segmentation by enforcing semantic consistency. Firstly, an attention module is added to the adversarial image translation process to get the most discriminative regions (ruler and background). Subsequently, a novel attention-consistent loss serves as an extra constraint to enforce the preservation of the attention-related information during the translation. Notably, we also propose a new saliency-consistent loss as another supervision to guarantee salient regions are consistent before and after translation. Then, we will use a well-trained convolutional neural network (CNN) model from the annotation-sufficient domain to achieve the segmentation task in another domain following image translation. The experimental results demonstrate the superiority of the proposed MSL model to solve the cross-domain segmentation problem for placenta photos. Furthermore, after the segmentation, an automatic pipeline is developed to extract specific visual characteristics from placenta photos for pathological diagnosis.Another topic is for studying the European art style known as realism. The specific case study we report here is the work of John Constable (1776-1837) who achieved a remarkable degree of naturalism in his sky painting. Specifically, we propose a new ML based paradigm for studying pictorial realism in a more objective way. Our framework assesses realism by measuring the similarity between two domains: clouds painted by artists known for their depictions of skies, like Constable, and photographs of clouds. The similarity is computed from two aspects, the accuracy of cloud classification and the discrepancy of painting styles. Our basic assumption is that if the paintings imitate observed reality well, painted clouds can be accurately classified by the classifier trained with cloud photos because they follow the similar feature representation documented in the photographs. The experimental results of cloud classification show that Constable more consistently approximates the formal features of actual clouds in his paintings than his contemporaries. We also put forward new evaluation metrics based on disentangled style features to carefully analyze the style discrepancy between paintings and photographs of clouds as well as between Constable and his contemporaries. We find that the painting style of Lionel Constable is most similar to the style of his father, John Constable, thereby verifying the opinions of art historians that he emulated his father's painting for practice. The study, as a novel interdisciplinary approach that combines machine learning, meteorology, and art history, is a springboard for broader and deeper analyses of pictorial realism.In addition, instead of only focusing on cross-domain artwork analysis, we discuss the potential biases that may appear in the general ML-based artwork analysis process as a supplement. Although recent research studies have substantially shown the effectiveness and applicability of machine learning for art understanding and creation, it has been argued that understanding or creating art is a human-defined procedure, which not only depends on stereotyped cognition or changeless laws, but also mixes with individual biases. Therefore, the procedure of analyzing artworks by virtue of intelligent machines can also be biased. Different kinds of biases such as labeling bias, design bias, and confounding bias can exist at any stage in the procedure of ML-based artwork analysis, unprofessional data curation, incomplete problem formulation, unreasonable algorithm design, subjective evaluation methods, etc. Since ignoring these biases could have negative impact on the precision and reliability of the experimental results, we are thus motivated to investigate these biases from both machine learning and art history perspectives and provide insights on how to systematically address them.

Subject Added Entry-Topical Term  

Artists.

Subject Added Entry-Topical Term  

Photographs.

Subject Added Entry-Topical Term  

Feces.

Subject Added Entry-Topical Term  

Maps.

Subject Added Entry-Topical Term  

Painters.

Subject Added Entry-Topical Term  

Landscape art.

Subject Added Entry-Topical Term  

Realism.

Subject Added Entry-Topical Term  

Painting.

Subject Added Entry-Topical Term  

Art history.

Subject Added Entry-Topical Term  

Art criticism.

Subject Added Entry-Topical Term  

Fine arts.

Index Term-Uncontrolled  

Visual data

Index Term-Uncontrolled  

Cross-domain image analysis

Index Term-Uncontrolled  

Image acquisition

Index Term-Uncontrolled  

Multi-region saliency-aware learning

Index Term-Uncontrolled  

Placenta segmentation

Added Entry-Corporate Name  

The Pennsylvania State University.

Host Item Entry  

Dissertations Abstracts International. 85-05A.

Host Item Entry  

Dissertation Abstract International

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:642896