Computer vision for structural damage quantification: A novel residual deep learning based approach

Recent advances of deep neural networks have revolutionized the techniques of machine learning for practical applications involving computer vision tasks. The flexibility of these models has allowed difficult tasks such as image segmentation to be tackled by this type of algorithms with much improved results. In this work, we propose and explore the capabilities of a novel deep residual neural networks with atrous convolutions for pixel to pixel classification tasks to achieve localization and quantification of structural damage in noisy image datasets. The proposed model is applied to a dataset of images synthesized to resemble debonding damage in honeycomb structures.