Specified Vacant House Image Analysis using Deep Metric Learning and Transfer Learning

Abstract

In recent years, the number of Specified Vacant House has been increasing due to the super-aging society. Specified Vacant House is a building in dangerous condition. Humans investigate whether Specified Vacant House or not. However, there is no objective criterion, and a unified judgment has not been made. To solve this problem, we introduce deep learning, which is good at image classification. We use a combination of deep metric learning and transfer learning methods to perform supervised learning on exterior walls. In this paper, we aim to obtain evidence to discriminate Specified Vacant House. As an experimental evaluation, we conducted AUC comparison for each learning model and feature visualization by GradCAM. As a result, the highest AUC (0.996) was obtained by transfer learning. In the visualization, we were able to focus on exterior wall features such as collapse and overgrowth.