Modeling and recognition of complex interference patterns in multi-surface phase-shifting interferometry

Traditional multi-surface phase-shifting interferometry (MS-PSI) algorithms are highly sensitive to phase-shift accuracy and experimental conditions. To address these limitations, we propose MSAI-Net (multi-scale attention interference network), which combines multi-scale convolutional feature extraction with SE attention and hierarchical LSTM-self-attention modules for robust fringe recognition. A simulated dataset incorporating nonlinear wavelength modulation and first- and second-order interference was constructed, with phase labels discretized into 500 classes. Simulation results show that MSAI-Net achieves 96.15% accuracy and remains robust to phase-shift errors and higher-order interference. Experimental validation further demonstrates low RMS error under phase deviations up to ±25%, confirming that MSAI-Net provides a reliable solution for high-accuracy multi-surface interferometric measurement.