Digital speckle pattern interferometry (DSPI) is a key technology in the field of precision non-contact measurement. It captures speckle interference fringe variations caused by surface deformation through machine vision combined with phase unwrapping algorithms to extract strain information, and is widely applied in non-destructive testing, precision measurement, and analysis of material mechanical properties. However, speckle images are susceptible to environmental noise, optical system distortions, and non-uniform detector sensitivity during actual measurements, which can lead to distortions in the phase unwrapping results, increasing measurement errors in speckle interferometry systems and seriously affecting non-contact measurement results.

To overcome the limitations of traditional phase unwrapping methods in terms of noise robustness, computational efficiency, and adaptability to discontinuous regions, and based on the deficiencies of current deep learning-based unwrapping algorithms, the UNet framework was adopted with targeted improvements. By integrating residual structures, attention mechanisms, and feature pyramid modules, the ResAtt-UNet speckle interferometry phase unwrapping algorithm was designed. This algorithm retained the classic U-shaped topology while innovatively introducing a bidirectional feature pyramid mechanism, constructing a symmetrical system composed of a feature downsampling path (encoder) and a feature upsampling path (decoder).

The algorithm was trained on simulated datasets and tested on simulated data. The results showed that the algorithm achieved a mean root mean square error (RMSE) of 0.0837 and a mean structural similarity index (SSIM) of 0.9996 (Figure 4), demonstrating the excellent phase unwrapping performance of the ResAtt-UNet algorithm. To validate the effectiveness of the UNet model improvements, ablation experiments were designed to analyze the influence of NAMAttention, SE-Model, and FPF-Convblock on model performance. The results showed that all three components effectively enhanced the UNet model’s performance, and the combination of the three exhibited significant synergistic effects (Table 1). Based on the principles of speckle interferometry deformation measurement, a speckle interferometry aluminum disk deformation measurement system was designed (Figure 8). The proposed algorithm was experimentally validated using this system. The results indicated that the maximum relative deformation calculated by the algorithm was 3.3800 µm (Figure 9), which was close to the true value. The unwrapped results were clearer and smoother, less influenced by background noise, demonstrating good robustness and generalization ability.

The ResAtt-UNet algorithm significantly improves the accuracy and efficiency of phase unwrapping in noisy environments, providing a reliable solution for dynamic deformation measurement.