Depth estimation from a single-shot fringe pattern based on DD-Inceptionv2-UNet.

Area of Science:

• Computer Vision
• Metrology
• Machine Learning

Background:

• Accurate 3D measurement from single-shot fringe patterns is crucial but challenging.
• Deep learning techniques are increasingly applied to fringe projection profilometry (FPP) for single-shot 3D measurement.
• Existing methods often face limitations in depth estimation accuracy.

Purpose of the Study:

• To propose an improved deep learning model for accurate depth estimation from single-shot fringe patterns.
• To enhance the accuracy of 3D measurement in fringe projection profilometry.
• To address the limitations of current deep learning approaches in single-shot FPP.

Main Methods:

• Development of the depthwise separable Dilation Inceptionv2-UNet (DD-Inceptionv2-UNet) model.
• Simultaneous adjustment of network depth and width for optimized performance.
• Evaluation using both simulated and experimental datasets.

Main Results:

• The proposed DD-Inceptionv2-UNet model demonstrated reduced error in depth map prediction compared to the baseline UNet.
• Mean Absolute Error (MAE) decreased by 35.22% on simulated data and 34.62% on experimental data.
• Qualitative and quantitative evaluations showed superior performance, with depth curves closely matching ground truth.

Conclusions:

• The DD-Inceptionv2-UNet model effectively improves the accuracy of 3D measurement from single-shot fringe patterns.
• The proposed method offers a significant advancement in depth estimation for fringe projection profilometry.
• The model shows strong potential for real-world applications requiring precise 3D reconstruction.