Identification of Features of Architectural Heritage Using Deep Learning Techniques

Abstract

Over the last decade, several changes have been witnessed at the Jeddah Historic Centre, comprising heritage buildings built during the medieval era. Identifying heritage buildings in historic cities is a complex, time-consuming process that requires the expertise of professionals and multiple verification steps. Technological advancements in artificial intelligence (AI) provide tools that can facilitate heritage identification. This study explores the application of the Deep Learning (DL) Convolutional Neural Networks (CNN) algorithm in identifying architectural features of heritage buildings in Al-Balad Area, Jeddah City, Saudi Arabia. To achieve this objective, a photographic survey covering most of the buildings in the area was conducted, and the architectural features of the images were analyzed through training (CNN) to identify their main features (Rawshan, windows, parapets, shops, and doors). The study results show that large data sets are not essential to train DL models as long as there is a balanced number of images across categories. Another affecting factor is the Image quality in terms of lighting, angle, clearly defined objects, carefully defined features, and labeling frames, which are essential for improved precise detection of Heritage-related elements. In conclusion, with the defined settings, the model could successfully identify architectural heritage features with an average precision of 88.2%, a recall rate of 4.3%, and probability thresholds of 54%. These results showed that Artificial Intelligence technology can contribute significantly to tangible heritage identification worldwide.