Evaluation of a pretrained deep learning model for indoor crack detection using DSLR and mobile phone cameras

Pretrained deep learning models have shown strong potential in automating crack detection for structural health monitoring. Most of these models are trained using datasets captured in outdoor environments under natural lighting. In addition, many crack detection models operate on two-dimensional images, which lack geometric context and limit the spatial interpretation of defects. The iTwin Capture Modeler by Bentley Systems addresses this limitation by integrating pretrained detection models with photogrammetric processing, enabling cracks to be detected and visualised directly on three-dimensional (3D) models. However, the pretrained model was developed using outdoor environments with image resolution of around 1 cm/pixel. Hence, this study aims to evaluate its performance under indoor conditions, where lighting and surface texture may differ significantly. Images were collected using a Digital Single Lens Reflex (DSLR) camera and a mobile phone. The DSLR produced native high-resolution images, whereas the mobile phone relied on pixel binning to improve image clarity in low-light situations. Both sets of images were used to generate 3D models through photogrammetric techniques, and crack detection was performed inside the iTwin software. The performance of the crack detection model was then evaluated by calculating its precision, recall, and F1-score. The DSLR camera recorded higher scores across all performance measures due to its superior optical quality and greater manual control. The mobile phone also provided satisfactory results despite having hardware limitations. These findings indicate that the pretrained model remains effective for detecting cracks in indoor environments and can be applied using a variety of image capture devices for three-dimensional inspection workflows.