Resolution-aware U-Net with state-space and attention modeling for speckle denoising of wrapped phase maps in digital holography

Speckle noise severely degrades wrapped phase maps in digital holography by distorting fringes and obscuring phase-discontinuity information that is essential for subsequent quantitative analysis. Existing CNN-based methods are limited by their small receptive fields, whereas transformer-based methods often incur high computational cost at high resolution. To address these issues, we propose a resolution-aware U-Net framework for wrapped-phase despeckling. The framework selectively applies global modeling strategies according to the feature resolution at different stages. Specifically, state-space modeling is used in shallow high-resolution layers to preserve long-range fringe continuity under strong speckle noise, whereas self-attention is introduced in deep low-resolution layers to enhance global phase consistency. To handle the inherent 2 π phase discontinuities, we adopt a sine-cosine representation that simplifies network learning. Extensive experiments on both simulated holographic data and real digital holographic measurements demonstrate the effectiveness of the proposed method. Under the most challenging mixed-noise condition, namely a severe combination of coherent speckle and additive Gaussian noise with a standard deviation of up to 0.4 rad, the proposed method reduces RMSE by approximately 38% relative to the Swin-Transformer-based baseline, demonstrating strong robustness and practical value for wrapped-phase denoising.