Data Augmentation and Normalization Optimization for Single-frame Fringe Projection Depth Estimation

This study aims to enhance the accuracy， robustness， and training stability of deep-learning-based single-frame fringe projection profilometry. It focuses on two practical limitations commonly encountered in existing methods： the incompatibility between conventional intensity normalization schemes and the highly skewed brightness distribution of fringe images， and the lack of physically consistent data augmentation strategies under limited labeled data conditions. The objective is to develop a preprocessing framework that preserves effective fringe information， improves model adaptability across different neural network architectures， and provides a model-aware optimization strategy without altering network structures or learning objectives.To achieve this goal， a unified preprocessing framework integrating Range-clipped Piecewise Contrast Normalization （RPCN） and Structured Stripe Occlusion Augmentation （SSOA） is proposed. RPCN is specifically developed to address the intrinsic characteristics of fringe images， which typically exhibit a large proportion of low-intensity pixels and a small number of extreme outliers caused by saturation， specular reflection， or sensor noise. The method combines interval clipping to suppress extreme pixel values， power-law mapping to enhance contrast in dark fringe regions， and piecewise remapping to separate shadow-dominated regions from effective fringe regions. This method redistributes pixel intensities into a range more suitable for neural network learning while maintaining stripe continuity and contrast stability. In parallel， SSOA is introduced to improve data diversity in a physically meaningful manner. Instead of directly masking or zeroing pixels， SSOA generates structured occlusions whose missing regions are reconstructed using neighboring stripe patterns， thereby preserving spatial continuity and phase-related information. This augmentation strategy simulates realistic measurement disturbances such as partial occlusion， surface defects， and local signal loss while maintaining physical consistency. The proposed framework is evaluated on multiple network architectures with varying capacities， including UNet， UNet++， and DeepLabv3+ with a MobileNetV2 backbone， to assess its generality， adaptability， and sensitivity to network complexity.Extensive experiments demonstrate that single-frame depth estimation performance is significantly improved across different network architectures by the proposed preprocessing strategies， while distinct effects are observed depending on model capacity. On the UNet architecture， the mean absolute error was reduced by 17.9% through the application of RPCN， indicating enhanced feature representation under limited normalization adaptability. When SSOA was applied independently， the root mean square error was reduced by 9.8%， and the peak signal-to-noise ratio was increased from 28.48 dB to 29.38 dB， reflecting improved robustness to structured disturbances. The joint application of RPCN and SSOA further increased the structural similarity index from 0.967 8 to 0.971 8. For the more powerful UNet++ architecture， the combined RPCN-SSOA framework yielded the best overall performance， with the structural similarity index increased from 0.977 2 to 0.979 1 and the peak signal-to-noise ratio improved from 29.73 dB to 30.34 dB. In contrast， for lightweight models such as DeepLabv3+ with a MobileNetV2 backbone， performance gains were constrained when SSOA was introduced， as the increased data complexity hindered effective learning， whereas a more favorable balance between accuracy improvement and training stability was achieved using RPCN alone. Additional evaluations across multiple test scenes further confirm that prediction stability is consistently improved under varying surface reflectance and illumination conditions. These results indicate that preprocessing strategies interact strongly with network capacity and architectural design， and that inappropriate augmentation may limit performance gains in compact models.The proposed RPCN-SSOA preprocessing framework provides a practical and adaptable solution for improving deep-learning-based single-frame fringe projection profilometry. By aligning intensity normalization with fringe-specific brightness distributions and enforcing physical consistency in data augmentation， the framework enhances depth estimation accuracy and robustness without increasing model complexity. The results highlight the importance of model-aware preprocessing demonstrate that normalization and augmentation strategies should be tailored to network capacity rather than uniformly applied. This work offers a feasible preprocessing paradigm and practical guidance for fringe-based three-dimensional measurement systems in both research and engineering applications.