CoMPD-Net: A collaborative multi-task deep learning framework for end-to-end phase demodulation of single-frame interferogram

Deep learning-based phase demodulation of single-frame interferograms shows great promise for online optical inspection. However, the large and fluctuating range of the unwrapped phase makes it difficult for neural networks to capture both the periodic and local detail features of the interferogram. Existing methods often rely on additional algorithms or large datasets, limiting their ability to achieve end-to-end, low-data-scale, high-precision demodulation. To address this, we propose a Collaborative Multi-task Phase Demodulation Network (CoMPD-Net). Guided by the principles of task decoupling, task collaboration, and physics-prior guidance, the network introduces a multi-task branch composed of classification and regression branches, which are designed to learn period-aware and local detail features, respectively. A bidirectional collaboration mechanism (BCM) is embedded within this multi-task structure to enable cross-task feature interaction and complementary learning. Meanwhile, a decision fusion module guided by physical priors refines the outputs via multi-scale median filtering and gradient-driven constraints, suppressing local phase anomalies. Furthermore, a physics-constrained loss system is developed to guide the model in learning phase distributions consistent with the physical model of optical interference. Simulation and experimental results demonstrate that CoMPD-Net achieves high- precision phase demodulation under a low-data-scale, without requiring any post-processing, thereby enabling true end-to-end phase demodulation.