Publications / 2018 Proceedings of the 35th ISARC, Berlin, Germany

Classification of Images from Construction Sites Using a Deep-Learning Algorithm

Daeyoung Gil, Ghang Lee and Kahyun Jeon
Pages 176-181 (2018 Proceedings of the 35th ISARC, Berlin, Germany, ISBN 978-3-00-060855-1, ISSN 2413-5844)
Abstract:

Field engineers take and collect several pictures from construction sites every day, and these pictures serve as records of a project. However, many of these images are loaded to and remain on computers in an unorganized manner because tagging, renaming, and organizing them is a time - consuming process. This paper proposes a method for automatically classifying construction photographs by job - type using a deep - learning algorithm. The first goal of this study is to classify construction images according to 27 job - types based on OmniClass Level 2. Google Inception v3 — a deep learning algorithm used in this study as an image classifier — was trained using 1,208 construction pictures labeled by job-type. To improve the performance of the classifier, the optimized number of training s was determined by examining the changes of accuracy and cross - entropy during trainings. The first result shows the incidence of several trainings over 50,000 was not meaningful. The re trained Google Inception as a construction image classifier was validated using a total of 235 images. The validation result shows that the classifier demonstrates an accuracy of 92.6% in classifying inputs properly and an average precision of 58.2% in correct classification. This means that retrained classifier can classify approximately nine out of every ten images correctly and that the deep-learning algorithm has high potential for use in the automatic classification of images from construction sites.

Keywords: Image classification, Deep learning, Construction site monitoring, Convolutional neural network,