DEGRADATION DEGREE ESTIMATION OF DISTRESS IMAGE IN ROAD INFRASTRUCTURE USING MULTI-DATASET CONTRASTIVE LEARNING

Abstract

For training a deep learning model that can estimate the degree of degradation from distress images in road infrastructures, pairs of distress images and their degradation degrees as labels are needed. Although a large number of pairs is desirable for achieving high estimation performance, the total number of such pairs is limited. On the other hand, there is an open dataset composed of distress images in other infrastructures. It is expected to improve the estimation performance of degradation degrees by using distress images of the open dataset in addition to the distress images of the road infrastructure dataset. However, in the open dataset, the distress images are not annotated with labels indicating the degradation degrees. Therefore, we propose a method for estimating degradation degrees across multiple datasets by introducing contrastive learning, regardless independent of the presence or absence of labels. The multi-dataset contrastive learning is performed as a pre-task of supervised learning. The obtained model parameters are used in supervised learning to estimate the degradation degrees of distress images in road infrastructures, and it is possible to achieve the improvement of estimation performance. The effectiveness of the proposed method is verified through experiments using real-world distress images in road infrastructure and an open dataset.