TILE DAMAGE DETECTION IN TEMPLE FACADE VIA CONVOLUTIONAL NEURAL NETWORKS

Many tiles on facade of historic temples are damaged and deteriorated due to aging and environmental factors. Regular inspection is required for proper maintenance and automatic inspection offers a significant advantage over manual inspection as it is efficient and accurate. Many previous studies are focus on detecting damages in factory tiles for the quality control purposes, although there has not been much study on tile damages for historical temples. This paper proposed an image-based system to detect damages in tiles using Convolutional Neural Networks (CNN) on a temple facade. The dataset was created by an Unmanned Aerial Vehicle (UAV) and a digital camera from a historical temple in Bangkok, Thailand. In the proposed work, CNN was trained on various image patches sizes. The detection accuracy of the system was found to be 95% in the validation data and 91% on the testing data. The results of the proposed system were compared with a system using hand-crafted features, including 2D wavelet transforms with the Artificial Neural Network (ANN). The proposed system shows that the CNN approach is more accurate than the traditional handcrafted method.