Evaluation of training data quality for deep learning-based damage detection

Tomotaka Fukuoka, Mai Yoshikura, Makoto Fujiu

Last modified: 2024-05-06

Abstract


A bridge inspection needs a lot of costs. It causes a lack of engineers and budget. So, some local governments couldn’t complete the bridge’s aggressive preventive maintenance in Japan. Recently, deep learning-based damage detection methods have been studied by many researchers to reduce the cost of the bridge’s aggressive preventive maintenance. This kind of method could detect the damage to the bridge by photo image with a detection model which has been trained with a large training dataset. In contrast to the increase in dataset size, the effectiveness of the quality of training data is not discussed enough. In this paper, the ratio of negative samples in the training data is regarded as the quality of training data.
In this study, we targeted the automatic detection process of the rebar exposure and the peeling on the surface of bridges. The effects of the ratio of the negative example data in a training dataset have been evaluated. In this study, the negative sample means the annotated data with no target damage. The negative sample image is the part of a bridge but does not include target damages. Many previous studies generate this kind of image when generating annotation data from real bridge photo images, but do not utilize it. A three-fold cross-validation method has been adopted to keep the robustness of the detection. The seven different detection models have trained with seven different training dataset which has different negative sample ratios. As a result of comparing the detection results of each model, the influence of the negative example data on the training data was confirmed. There was a peak of the recall curve in the middle of increasing the negative sample ratios. The correct negative sample ratio could improve the accuracy of damage detection.

Keywords


image processing; damage detection; noise ratio

Full Text: PDF