ADDITIONAL LEARNING OF SEGUMENTATION MODELS USING BACKGROUND IMAGES OF INSPECTED BRIDGES

Abstract

Visual inspection is important for the maintenance of bridges. However, the decrease in the working population in the construction industry has become an issue in Japan. In addition, visual inspection is time consuming and dangerous in some cases. Therefore, the efficiency, rationalization, and safety of inspection work are required. The current inspection content need decision making based on the experience of the inspector. This can lead to serious accidents due to human mistakes. Inspection methods that utilize AI and UAV can solve these problems. In this study, we performed automatic damage detection of UAV images using the U-Net model. The balance of the dataset was ensured by focusing only on corrosion. The problem that UAV images have a large proportion of background and are prone to false positives was improved by background reinforced training. This method is to train a U-Net or other semantic segmentation models by standard damage annotated image data, and training it again before use it to real bridge UAV videos by a few background non annotated images to let the background looked familiar to the model. The background reinforced training of UAV images resulted in improved detection accuracy. It is considered that this is because the model learned the characteristics of the bridge and the information around the bridge from the UAV image.