RGB and Thermal Image Analysis for Marble Crack Detection with Deep Learning

Surface damage identification implies visual inspection, which is traditionally performed manually with the bare eye. The latter is a time-consuming and error-prone process. Effectively automated inspection systems would provide sustainable solutions. The aim of this work is to investigate the development of an automated inspection system based on non-contact thermal and color imaging for detecting cracks on marble slab surfaces. Marble crack detection is challenging due to the nature of cracks being invisible to the human eye due to their fine thickness, lost in the randomly textured marble surfaces. In this work, a comparative performance evaluation is conducted between several state-of-the-art deep learning (DL) models, using both thermal and red–green–blue (RGB) marble slab images. More specifically, four DL semantic segmentation models and 28 feature extraction backbone networks are combined to form a set of 112 modular DL architectures. Results indicate the best-performing model architecture, the FPN, reporting 71.61% and 68.07% mean intersection over union (mIoU), for RGB and thermal images, respectively, while best-performing backbone, the efficientnetb4, reporting 80.07% and 75.49% mIoU for RGB and thermal images, respectively.