A deep learning based facade analysis method for detection of periodic differences in housings: the case of Konya

Architectural heritage, which is an important component of cultural sustainability and urban identity; It is very important to be able to observe the sociological, economic and technological accumulations of the past societies and to benefit from these accumulations today. Categorizing the architectural accumulations of past societies contributes to better understanding of the characteristics of civilizations. In the analysis of the period and style characteristics of architectural heritages; archeology, art and architectural history etc. experts from different disciplines work together. These analyzes on architectural heritage are shaped by the experience of experts and the methods they use. As a result of the literature studies, it is seen that artificial intelligence methods, which have become popular in recent years, make significant contributions to experts in architectural heritage studies. However, there is no study in which artificial intelligence-based deep learning methods are used in determining the periods and styles of architectural heritages. In this thesis study; The usability of deep learning-based object detection methods as an alternative tool has been tested in order to contribute to the ability of experts to perform more practical and sensitive analyzes in period and style detection studies on architectural heritage. Within the scope of the study, an algorithm has been developed that can perform period detection by analyzing the facade images of traditional Konya houses. In order to make the proposed algorithm applicable, first of all, 448 images of 80 traditional houses built in 4 different periods in the historical city center of Konya were compiled as a result of field study. 400 of the compiled image were labeled according to their periods with the LabelMe application and used in the creation of the training and validation data set of the developed algorithm. In the algorithm, which is trained on the images in the datasets, is used Faster R-CNN architecture with a fast and high accuracy capacity. Finally, 48 unlabeled images belonging to 4 different periods were tested with the algorithm that successfully completed the training. According to the findings obtained from the study; it has been observed that the algorithm used to determine the periods of traditional Konya houses performs period detection with an accuracy rate of 96% and above on the images. This shows that the developed algorithm will be a helpful method for experts working on architectural heritage. As a result of the thesis study; not only contributed to the literature on traditional Konya houses, but also proposed a new method that can be used in the period and style determination of architectural heritages in the future.