Effectiveness of Background Removal for Corrosion Classification of Tunnel Lighting Fixtures Using Transfer Learning with ResNet18

Abstract

In the field of social infrastructure, equipment inspections are still predominantly carried out through visual observation. In recent years, image classification techniques based on Convolutional Neural Networks (CNNs) have garnered significant attention for the purpose of automating this process. However, it is often challenging to obtain sufficient data from real-world sites. Therefore, transfer learning using models pretrained on large-scale datasets has become a common approach. In transfer learning, convolutional layers are typically frozen, which may lead the model to respond to irrelevant background regions in the image. Therefore, background removal is expected to reduce the model’s sensitivity to irrelevant background regions. In this study, we examined the impact of background removal on the classification performance of transfer learning models applied to images of mounting hardware for lighting fixtures in highway tunnels. The results suggest that background removal may contribute to improving the classification performance of these models. for performance improvement in models with constrained feature extraction capability.