Semantic segmentation of cracks: Data challenges and architecture

Deep Learning (DL) semantic image segmentation is a technique used in several fields of research. The present paper analyses crack as case study to review the up date research on presence fine structures and effectiveness established approaches address inherent class imbalance issue. UNet architecture tested against networks consisting exclusively stacked convolution without pooling layers (straight networks), with regard resolution their results. Dice Focal losses are also compared each other evaluate highly imbalanced data. With same aim, dropout data augmentation tested, additional regularizing mechanisms, uneven distribution dataset. experiments show that good selection loss function has more impact handling boosting detection performance than all regularizers regards resolution. Moreover, UNet, considered reference, clearly outperforms no both training time. authors argue architectures, layers, achieve high at very low cost terms Therefore, consider such state art for cracks. On hand, once computational not an issue anymore thanks constant improvements technology, application might become attractive again because simplicity performance.