Automated Masonry crack detection with Faster R-CNN

Inspection of masonry buildings, typically railway bridges, for crack detection is currently performed by humans under tedious and sometimes dangerous working conditions. Over the past years, computer vision based techniques have been developed to automate structure visual inspections. These techniques could be integrated with (semi) autonomous drone surveillance to collect images of assets for full automation of simultaneous inspection and crack detection in railway bridges. In this study we have adopted the architecture of Faster R-CNN object detectors to provide crack detection in images. In this architecture, we have tested three networks (Mobilenetv2, Resnet50 and ZF512) to be utilised as feature extractors in a limited resource system for crack detection. We propose a new way of performing detection that we call Progressive Detection to increase the robustness of detection, considering otherwise only partially detected cracks. Since one of the main goals of visual inspection is checking the health of every single defect, we have revisited binary classification of images with and without cracks from a detection point of view, with the objective of minimising crack missing rates. Results show that Mobilenetv2 performs both successfully and fast enough to be applied in a drone application as a feature extractor network, achieving a close level of performance to the more sophisticated network Resnet50 with half its inference time. Regarding classification, Mobilenetv2 achieves its best performance in the early stages of its training process, showcasing 93% accuracy and a crack miss rate ranging from 1% to 15%. These results are comparable to Resnet's and better than ZF512's.