Research on Structured Light Wrapped Phase Unwrapping Algorithm Based on Deep Learning

Phase unwrapping serves as a critical step in high-precision optical measurement， aiming to recover the absolute phase from its wrapped counterpart. While current deep learning-based approaches for absolute phase extraction have achieved improvements in both accuracy and computational efficiency， critical challenges persist， including excessively large network parameter counts and low accuracy in resolving discontinuous phases. To overcome these limitations， this paper introduces the Attentive Directional Encoder-Decoder Network （ADE-Net） for structured light phase unwrapping applications. The primary objective is to develop a lightweight， efficient， and robust deep learning framework that achieves consistently high unwrapping accuracy under complex conditions such as noise corruption， phase discontinuities， and aliasing artefacts， thereby providing a practical， engineering-ready solution capable for reliable development in real-world optical measurement systems.The proposed ADE-Net is built upon a concise and efficient encoder-decoder backbone architecture. The encoder consists of four feature extraction stages， each comprising a convolutional block and a Haar Wavelet Down-sampling （HWD） module. The HWD structure preserves high-frequency directional details， such as phase jump edges， during down-sampling， providing a more complete feature foundation for subsequent processing. The decoder employs a symmetric up-sampling structure based on transposed convolutions， with skip connections to fuse multi-scale features from the encoder， ensuring effective detail recovery. To enhance feature representation， two innovative modules are integrated： the Adaptive Clockwork Long Short-Term Memory （AC-LSTM） module and the Depthwise-separable-convolution Efficient Multi-Scale Attention （DEMA） module. The AC-LSTM module is designed to capture multi-scale contextual and direction-sensitive features simultaneously. It consists of two parallel branches： an Atrous Spatial Pyramid Pooling （ASPP） branch for multi-scale context aggregation， and a bidirectional LSTM （Bi-LSTM） branch enhanced with Context Anchor Attention （CAA） to model long-range dependencies along horizontal and vertical directions with spatial awareness. The outputs of both branches are fused to produce rich， direction-aware features. The DEMA module is an improved version of the Efficient Multi-Scale Attention （EMA） mechanism. It introduces depthwise separable convolution to replace standard 3×3 convolutions， significantly reducing computational complexity while maintaining receptive field coverage. Additionally， Batch Matrix Multiply （BMM） is adopted for more efficient parallel processing of batch data. DEMA operates by grouping input features， performing adaptive height and width pooling， and applying depthwise separable convolution in parallel branches， followed by cross-branch fusion via BMM and element-wise multiplication with the original features. This design enhances the model's ability to focus on critical information across scales while optimizing computational efficiency.To comprehensively evaluate the performance of ADE-Net， systematic experiments were conducted using both simulated and real-world datasets. For simulation， three distinct datasets were constructed， each comprising 2 000 pairs of wrapped and absolute phase images with a size of 256×256. Dataset 1 simulated noise interference by introducing Gaussian noise with signal-to-noise ratios of 0， 5， 20， 40， and 80 dB. Dataset 2 simulated phase discontinuities by randomly generating 1 to 2 rectangular jump regions with random positions， sizes， and phase values. Dataset 3 combined both noise and discontinuities to emulate aliasing effects. Additionally， the public real dataset ‘Single-input dual-output 3D shape reconstruction’ was adopted to assess generalization capability， which contains 1 500 real-scene samples acquired by a structured-light 3D measurement system. The evaluation was performed using three metrics： Normalized Root Mean Square Error （NRMSE）， Mean Absolute Error （MAE）， and Structural Similarity Index Measure （SSIM）. Benchmark comparisons were carried out against mainstream models such as U-Net， Res-UNet， Perera's Net， and TNUNet. In the comparative experiments， ADE-Net achieved the best performance across all three simulated datasets. On the discontinuous dataset， it attained an NRMSE of 4.35%， an MAE of 1.094 5 rad， and an SSIM of 0.871 4. On the random-noise dataset， the results were 4.58% NRMSE， 1.216 5 rad MAE， and 0.838 9 SSIM. On the aliased dataset， ADE-Net reached 4.34% NRMSE， 1.137 7 rad MAE， and 0.834 0 SSIM. Cross-validation on the public dataset demonstrated excellent generalization ability， with a Root Mean Square Error （RMSE） of 0.505 2， an MAE of 0.180 9 rad， and an SSIM of 0.997 1. Ablation studies further confirmed the individual performance contributions of both the AC-LSTM and DEMA modules. Moreover， while maintaining the aforementioned high accuracy， ADE-Net also exhibited remarkable computational efficiency advantages. With only 2.012 M parameters and 2.655 GFLOPs， it is significantly lighter than the compared models. The inference time per single image is only 12.71 ms， meeting real-time processing requirements without compromising accuracy， thereby achieving an effective balance between model complexity and computational performance.In conclusion， ADE-Net presents a lightweight， efficient， and highly accurate deep learning framework for phase unwrapping in structured light 3D measurement. By integrating the AC-LSTM module for capturing direction-sensitive multiscale features and the DEMA module for computationally streamlined attention-driven feature refinement， the models achieve an ideal balance between performance and computational complexity. It demonstrates superior resilience in challenging conditions involving noise corruption， phase discontinuities， and aliasing artefacts， outperforming leading benchmark models in both accuracy and processing efficiency. Rigorous validation on real-world datasets confirms its robust generalization capability， while its fast inference speed and low resource footprint make it well-suited for high-precision phase reconstruction tasks. By addressing practical constrains in deployment， this work promotes the feasibility of structured light 3D measurement system for real-world， resource-aware applications.