Automatic Thin Crack Segmentation with Deep Context Aggregation Network

Automatic crack segmentation in infrastructure is an important and challenging topic in the field of computer vision. Cracks on the surface of some infrastructure such as nuclear reactor containment are usually thin, and most segmentation methods lose a lot of features due to the use of downsamplings. Inspired by the small object segmentation methods, we propose a crack segmentation network named UCAN based on the effective application of dilated convolution. UCAN can extract more contextual features of thin cracks while preserving resolution by dilated convolutional layers with an exponentially continuously changing dilated rates. In addition, we propose a new loss function that can pay more attention to miss detection in segmentation loss. Finally, we experimentally verify that the proposed method can obtain the best F1-Score on the thinner datasets.