Robust single-shot fringe analysis via deep learning for extended depth

To address the depth-of-field (DOF) limitation caused by defocus in structured light 3D measurement, this paper proposes a deep learning-based single-frame defocused fringe phase extraction method. The approach enables high- precision phase recovery across an extended DOF range using only a single acquisition. Traditional methods, which rely on multifocal plane fusion or complex optical designs for DOF expansion, suffer from low efficiency and system complexity. The proposed method introduces a two-stage convolutional neural network (CNN) framework: the first stage, a defocus compensation network, takes a single defocused fringe image as input and jointly generates a refocused sharp fringe pattern and modulation map through multiscale residual learning, effectively mitigating contrast degradation from defocus blur. The second stage, a phase decoding network, employs a frequency-guided attention mechanism to directly predict the numerator and denominator terms of the wrapped phase from the first- stage outputs, enabling absolute phase reconstruction via arctangent calculation. To enhance generalization, a physics-driven synthetic data generation method was developed to simulate fringe patterns under varying defocus levels and noise conditions, complemented by a transfer learning strategy for real-world adaptation. Experimental results demonstrate that the method achieves an average phase extraction error of 0.02 rad when extending DOF to three times that of conventional approaches. By requiring only single-frame acquisition, the system significantly improves measurement efficiency, particularly in dynamic scenes or resource-constrained environments. This work provides a novel solution for single-frame extended-DOF structured light measurement, with promising applications in industrial inspection, biomedical imaging, and related fields.