When serious earthquakes occur, people in the affected area often suffer the several difficulties such as victims by aftershocks, long-term evacuation, and reconstruction of their households. These difficulties were caused by various factors, especially delay of current investigation procedure of earthquake-damaged houses. In fact, in The Kumamoto Earthquake in 2016 and in Great East Japan Earthquake in 2011, investigators couldn’t finish their works easily and quickly because of lack of inspectors, the several aftershocks and floods from tsunami.
In order to solve these issues, several researches have been conducted. Particularly image processing with deep learning has been evolved because the image recognition has remarkably developed in recent years. In terms of timber houses, image classification is currently widely used, but it is difficult to adapt these methodologies into actual investigations because image classification is good at qualitative damage assessment rather than quantitative ones.
Then, object detection and image segmentation have been attracting attention. This is because they have a potential for their automatic and speedy processing and enabling quantitative damage assessment. Actually, they were already used for infrastructures damage evaluation. In this study, the same technique is applied to the guideline of the image diagnosis of timber houses damaged by earthquakes.
Firstly, a damage extractor was created and verified. In order to create damage extractor, which is based on semantic segmentation, a database was made from a large number of photos from past big earthquakes. Then, the tagging all images in the database into four types of damage was carried out, in order to assess these types. However, this database didn’t work well at the step of deep learning because of the lack of images and the bias of image feature values. Thus, chromakeying is employed and it enabled to improve deep learning accuracy, and the effectiveness of chromakeying to the deep learning database was also confirmed.
Secondly, damages were extracted, and its rate were calculated using only deep learning image processing. There were various noises in extract results which means difficulty of quantitative damage assessment. Then in order to reduce these noises, pre- and post-image processing, accuracy of damage extraction was improved. The quantitative damage assessment based on the guideline of damage assessment for earthquake insurance, was conducted using damage extractors, damage rate of some samples of mortar walls was calculated.
Finally, for the improvement of image diagnosis, not only surface damage but also structural parameters should be considered, and the correlation between them was focused on. Using the previous proposed correlation, drift ratio would be estimated from surface damage. However, although based on previous researches and experiments the correlation has become clear, this is not enough for image diagnosis. Thus, the new experiment data was added to make the data improved. As a result, image diagnosis can roughly estimate the structural maximum drift ratio.
