|aImproving Deep Learning-Based Facade Visual Inspection :|bA Data Quality Perspective.
264
0
|c2021
300
|a1 online resource (335 pages)
336
|atext|btxt|2rdacontent
337
|acomputer|bc|2rdamedia
338
|aonline resource|bcr|2rdacarrier
500
|aSource: Dissertations Abstracts International, Volume: 84-04, Section: B.
500
|aAdvisor: Qian, Wang.
502
|aThesis (Ph.D.)--National University of Singapore (Singapore), 2021.
504
|aIncludes bibliographical references
520
|aA building facade is an external structure to support and protect a building. Defects occurring on facade can reduce the service life of the whole building. If timely inspection and maintenance are not conducted, facade defects may cause great damages to the surroundings. Therefore, facade inspection has drawn great attention in both industry and academia.Visual inspection that assesses the facade condition is the first and a key step for periodic facade maintenance. However, traditional approaches to achieving the facade visual inspection are criticized as risky, laborious, time-consuming, and expensive. To avoid the problems in the traditional approaches, there are increasing efforts exploring to use mobile devices (i.e., unmanned aerial vehicles and robots) and artificial intelligence techniques to inspect the facade condition. Particularly, deep learning algorithms have attracted considerable research interests in recent years for structural health monitoring. Although attempts have been made to achieve defects identification with different deep learning models for various structures, many problems still existed that weaken the performance of deep learning-based facade visual inspection. Especially, the problems caused by data quality were not well addressed in previous studies.The main objective of this thesis was to improve the performance of deep learning-based facade visual inspection. In this thesis, the performance was considered from the aspects of reliability and efficiency. To achieve this objective, this thesis designed a research methodology based on the theory of total data quality management. The methodology included four phases: definition, assessment, analysis, and improvement.In the first phase, the requirement of data quality was defined. Meanwhile, the critical procedures, activities, and human factors were extracted using a Delphi study. The results of the Delphi study revealed that the procedures of data pre-processing and model construction have greater influences on the uncertainty of efficiency and the uncertainty of reliability. Next, four data quality problems, including imbalanced distribution, incomplete information, ineffective labels, and inconsistent labels, were identified in the second phase of assessment. Then, in the third phase of analysis, the potential solutions to the four data quality problems were analyzed based on the idea of managing the existent data without burdening humans. In the last phase of improvement, a research framework consisting of the corresponding solutions was proposed. The research framework was unfolded through three procedures: data selection, data annotation, and model training. For each procedure, criteria were designed to assess the data quality, and solutions were developed enabling the target stage to focus on "better" data. For data selection, a semi-supervised learning solution and an active learning solution were designed by integrating the degree of uncertainty and the degree of representativeness to utilize the unlabeled data and to enrich the labeled data. For data annotation, annotation rules for categorization, localization, and segmentation were defined in accordance with the assessment standards for facade inspection. For model training, a meta learning method was applied for solving the problem of imbalanced distribution while a rule-based deep learning method was designed to regulate the learning direction. These two methods were combined with a criterion of the similarity to the ground-truth data to provide higher weights to the highquality data. The experiment results demonstrated that the proposed solutions successfully improved the accuracy and stability of the detection of various facade defects. Besides, the detection results obtained by the proposed solutions provided more effective outcomes for condition evaluation. Meanwhile, the time and cost were saved in general perspective because no extra labor works were expended. Therefore, the reliability and the efficiency of deep learning-based facade visual inspection were effectively improved by the proposed research framework.
借閱次數
: 
點閱：12
評分：0
評論：0
引用：0
