Detection of hidden defects in reinforced concrete based on DSSPI and a multi-scale multi-branch residual convolutional network

• Introduces a new phase acquisition method and dataset. • Proposes an efficient one-step phase unwrapping method. • Multi-dimensional analysis of the robustness of unwrapping. • Detection of hidden debonding defects in reinforced concrete. DSSPI metrology and phase unwrapping technology are commonly employed to detect hidden defects in objects and to assess strains and stresses at defects, making them highly significant for structural health monitoring. However, traditional methods to obtain the unwrapped DSSPI phase typically involve multiple steps, such as phase filtering and unwrapping, which reduces detection efficiency. To address this, we propose a residual convolutional neural network with a multi-scale and multi-branch architecture that enables one-step unwrapping of noisy wrapped phases, significantly improving detection efficiency. By mixing the real and simulated phases as input, the network’s generalization ability is enhanced. The RMSE between the unwrapped and label phases is calculated and compared against four traditional unwrapping methods to verify the robustness and stability of the proposed method. Results demonstrate that our method outperforms the other four methods in terms of both unwrapping accuracy and stability. When applied to the unwrapping of hidden defect phases in reinforced concrete, our method shows fewer discontinuities and better phase continuity. And compared with the label phase, the average pixel error only accounts for 2.8% (0-255) of the pixel range. Additionally, unwrapping efficiency is increased by 24.6 times, showing broad application prospects.