Deterioration Level Estimation for Infrastructures Considering Noisy Labels via DivideMix

This paper proposes deterioration level estimation for infrastructures considering noisy labels via semi-supervised learning. Although deep neural networks (DNNs) are often used for infrastructure maintenance recently, conventional DNNs suffer from noisy labels included in the distress image dataset caused by the annotations from multiple engineers with different skill levels. A recent study reports that DNNs are likely to overfit such noisy labels, which greatly aggravates the generalization performance. Therefore, this paper focuses on reduction in the negative impact of noisy labels via DivideMix, which can learn from noisy data through a semi-supervised learning approach. As a result, the proposed method enables the improvement of the performance in the deterioration level estimation by considering noisy labels. Experimental results obtained by applying our method to the actual infrastructure inspection data show the effectiveness of our method.