Deep Learning-based Semantic Segmentation for Crack Detection on Marbles

The demands for the improvement of production cost and quality have been increased with the widespread application of marble. The problem of detecting fractures and cracks on marble has attracted an increasing amount of attention recently. In this study, an artificial intelligence-based fracture and crack detection system with high accuracy is proposed. In the proposed system, images in the dataset were first trained with the convolutional neural network-based semantic segmentation model, DeepLabv3+, in order to detect fractures and cracks. The dataset has been augmented by adding different ground images containing fracture and crack surfaces as well as marble images due to the positive contribution of the size of the dataset to the accuracy. After training with 5 different convolutional neural networks and 3 different optimization algorithms, the mIoU and weighted mIoU values were achieved as 0.672 and 0.967, respectively, with the RMSprop algorithm and ResNet-50 architecture. Unlike similar studies, the proposed system has been integrated with our custom-designed interface. Thus, it is aimed to make fast and efficient crack detection on the production line.