Deployment Conscious Automatic Surface Crack Detection

Abstract

Automatic crack detection from images is a challenging problem long studied by many, and it has a potential to drastically reduce the labor intensive building and road inspections currently done manually. In this paper, we propose a deep learning based crack detection model with a strong emphasis on deployability. By applying various network architecture designs utilized in state of the art semantic segmentation models, as well as augmenting the network with Multiple Instance Learning (MIL), our model is able to accurately predict cracks at a significantly reduced inference time. Moreover, in order to realize a more holistic inspection automation, we propose a semi-supervised training method which enables a model to predict crack orientations in addition to predicting crack locations. Experimental results show that the proposed model surpasses the state of the art in terms of accuracy at a significantly reduced inference time.