Single-frame fringe analysis drops the hand-tuned filter window: spectral attention adapts per image, and simulation-trained models reconstruct unseen experimental data

Phase demodulation of single-frame fringe patterns moves from manually set Fourier-domain filter windows to a network that decides, for each input, which +1st-order spectral content to keep. Earlier spatial phase demodulation methods, including Fourier-transform analysis, depend on handcrafted analysis parameters such as the shape of the frequency-domain filter window, and suitable settings are hard to determine when fringes are severely degraded by noise, defocus or overexposure. On 400 simulated test samples the network leads on all four metrics, MAE, RMSE, PSNR and SSIM, with the smallest standard deviation, and against the next-best ResUNet it reduces MAE and RMSE by roughly 39% and 10.6%.

The network extracts and modulates features in the spatial domain, transforms them to the Fourier domain for adaptive spectral filtering, then returns to the spatial domain to regress the real and imaginary parts of a complex distribution and compute the wrapped phase. Spectral selection is carried by an adaptive spectrum filtering block that fuses spectral features with convolution and residual dense blocks, then uses a Sigmoid to produce attention weights between 0 and 1 that enhance +1st-order components and suppress irrelevant ones, formally matching the filtering step of Fourier-transform analysis. Training pairs come from a forward model of off-axis digital holographic microtomography that simulates a fiber's three-dimensional refractive index and adds noise and overexposure, giving 4400 pairs split 10:1 into training and validation; experimental interferograms come from a 532 nm laser system with a 20x/0.40 NA objective and are cropped to 512x512 pixels.

A spectrum attention mask is added during training, combining an inverted Hanning window that suppresses the central zero-order region with a half-plane mask that suppresses the conjugate spectrum, constraining where the attention weights search. Without the mask, the loss rises substantially in early training and stability drops, and repeated runs under the same configuration consistently diverged after about 10 epochs. The mask is multiplied with the attention weights before filtering and is required for training stability; the ablation also shows that removing spatial feature modulation hurts demodulation accuracy most, while removing the adaptive spectrum filtering block degrades performance less.

The method targets single-frame fringe pattern phase demodulation in digital holographic interferometry and is applied to three-dimensional refractive index reconstruction from 180 angular projection phases in digital holographic tomography. Simulation validation covers 400 test samples, and experimental validation uses fringe patterns recorded by a 532 nm laser tomography system with a 20x/0.40 NA objective and cropped to 512x512 pixels. Experimental interferograms have no precise ground truth, so the experimental part is a qualitative comparison against hand-tuned Fourier-transform results.

Although noise and overexposure are added to the simulated data, a domain gap remains between simulated and real fringe patterns, arising from coupled degradations such as nonuniform illumination, environmental and system noise, defocus blur, local saturation and mechanical vibration, which limits generalization on experimental data. Experimental interferograms lack precise ground truth, so quantitative evaluation is infeasible. Future work could consider domain randomization, self-supervised or unsupervised training on unlabeled experimental interferograms, and domain adaptation to narrow the simulation-to-experiment difference, with constructing reliable pseudo-labels still an open problem.