Subsurface defect area quantification of reinforced concrete structures with array ultrasound and dual-scale neural network

Array ultrasound is effective in detecting subsurface defects of reinforced concrete (RC) structures. However, the current practice of ultrasonic image interpretation remains manual and qualitative, restricting the automatic and intelligent subsurface defect quantification. This study proposes a subsurface defect area quantification method for RC structures with array ultrasound and dual-scale high-resolution neural network., Parallel high-resolution convolution streams and multi-resolution fusions were developed in the high-resolution network to generate spatially precise and semantically strong representations of defects. Dual-scale architecture was proposed based on the high-resolution network, taking advantage of global-scale context to assist local-scale network, and expecting to improve inference accuracy. RC specimens with multiple types preset artificial defects were designed and manufactured. C-scan images were acquired using total focus imaging method and low-frequency ultrasonic array and employed to train the dual-scale network. Individual plane maps output by dual-scale network were registered to global plane representation maps, and defect areas were quantified. Results reported that different types of defects can be distinguished from other high-intensity reflections in C-scans by the proposed deep learning model. Mean F-score and IoU of testing set were 88.50 % and 80.05 % respectively, and defect F-score and IoU were 86.24 % and 75.81 % respectively, all higher than the local-scale high-resolution network, demonstrating the superiority of dual-scale architecture. MAPE and R² of defect area quantification were 6.07 % and 0.9779, indicating the proposed method facilitates subsurface defect quantification to mm-level with high precision.