Crack detection with deep learning: an exemplary study of data design in architecture

---

Tezin Türü: Doktora

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mimarlık Fakültesi, Mimarlık Bölümü, Türkiye

Tezin Onay Tarihi: 2018

Öğrenci: ÇAĞLAR FIRAT ÖZGENEL

Danışman: ARZU SORGUÇ

Özet:

Dramatic increase of available data in the last 20 years transformed the role of data in artificial intelligence algorithms for problem solving. Deep learning embodies potentials for both finding novel correlations within data, and improvement in decision making process in its massiveness. Thus, this approach is prominent in processing such massive data by removing the necessity of explicitly determining features relevant to the solution. Reformulation of the problem in terms of determining which data represent the problem and evaluating the results emerge as the primary challenges in deep learning applications. Within the scope of this thesis, data design term is introduced to describe end to end process of problem solving with deep learning algorithms which is suitable for broad range of applications including problems in architecture. Data design defined as a holistic approach embracing the process from problem (re)formulation to evaluation of the results considering the interrelations of decisions made throughout the process. In this context, data design in architecture is exemplified with the task of crack detection in buildings in order to minimize subjectivity in the course of evaluating the results. For this purpose, the relation between data and deep learning framework, case specific evaluation requirements and strategies for enhancing the performance are inspected v vi through image classification and semantic segmentation applications for crack detection. Concordantly, this study contributes to the literature not only with the introduction and framing of data design but also with the proposal of crack detection specific evaluation metrics for both image classification and segmentation applications and a novel method is proposed employing quad tree and deep learning algorithms in conjunction for semantic segmentation of objects with limited visual features. As a result, data design and respective consequences are discussed in depth and demonstrated regarding the case dependency, decisions taken in the course of implementation and their influences to both process and the results.