深層学習に基づく物体検出と生成モデルを用いた構造表面損傷の実時間検出技術の開発

抄録

Recently, assessing the integrity of the structures accurately and reliably has become extremely important in various fields in order to increase operational lifetime and improve safety. Detecting surface damage such as cracks, spalling and so on in evaluating the soundness of the structure is particularly important as it is one of the major factors causing deterioration and destruction of the structure. In this study, we develop a real-time screening system for structural surface damage by using object detection technique and generative model based on deep learning. The accuracy of object detection depends largely on the quality and quantity of images given as training data in advance. These images are not kept even by road management companies, etc., and tend to be short. In this study, we generate images of the crack using deep generative model and introduce them into object detection technique. Finally, we investigate whether the detection accuracy of damage is improved by introducing generated images as training data.