Two-Stream Boundary-Aware Neural Network for Concrete Crack Segmentation and Quantification

Cracks can be important performance indicators for determining damage processes in new and existing concrete structures. In recent years, deep convolutional neural networks (CNNs) have shown great potential automatic crack detection segmentation. However, most of the current CNNs tend to lose high-resolution details and, therefore, lead blurry object boundaries; this results poor images with complex backgrounds engineering This study proposes a two-stream boundary-aware segmentation (BACS) network that combines semantic image semantically informed edge explicitly. Firstly, (HRNet) is utilized branch strong representations through repeatedly conducting multi-scale fusions across parallel convolutions. Furthermore, an preserving fine-grained elongated thin cracks, which adopts modified dynamic feature fusion (DFF) produce more accurate sharper predictions. The proposed method evaluated using dataset 1,892 three different scenarios. show mean intersection-over-union (mIoU) scores reach 79.26%, 68.74%, 70.31% pure crack, background, variable-width scenarios, respectively. addition, width quantification performed validate accuracy terms practice. BACS achieves high average absolute error 0.0992 mm, corresponds approximately two pixels images. conclusion, provides effective solution task, especially scenario, providing data foundation digital twin