Abstract
It is critical for infrastructure manager to keep the status high-quality for providing the service to users at daily activities. Using surveillance cameras and drone inspection toward the damage feature, there has been progress to automate its inspection toward the grade of health condition whether the deterioration has been changed or not. When we prepare a pair of raw images and damage class labels, it is possible to train supervised learning toward the predefined damage grade, displacement. However, such a damage representation does not always match the predefined classes of damage grade, so there may be some detailed clusters from unseen damage space, or more complex clusters from overlapped space between two damage grades. The damage representation has fundamentally complex feature, so all the damage classes could not be perfectly predefined. Our proposed MN-pair contrastive learning method enable to explore the embedding damage representation beyond the predefined classes including more detailed clusters. This method intends to maximize the similarity of M-1 positive images close to the anchor, and simultaneously to maximize the dissimilarity N-1 negative ones far apart, using both weighting loss function. This MN-pair method has been faster learning than the N-pair algorithm, instead of using one positive image. We propose a pipeline to learn the damage representation and to automate to discriminate more detailed clusters using the density based clustering on the embedding 2-D reduction space. We also visualize the explanation of damage feature using Grad-CAM for MN-pair damage metric learning. We demonstrate our method to three experimental studies such as steel product defect, concrete crack of deck and pavement, and sewer pipe defect. Furthermore, we mention the usefulness of our method and future works to tackle.
Abstract (translated)
URL
https://arxiv.org/abs/2301.06077