LiteF2DNet: a lightweight learning framework for 3D reconstruction using fringe projection profilometry.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Fringe projection profilometry (FPP) is a primary method for 3D object profiling.
• Traditional FPP algorithms suffer from error propagation due to multistage processing.
• Deep learning offers end-to-end solutions for more faithful 3D reconstruction.

Purpose of the Study:

• To introduce LiteF2DNet, a lightweight deep learning framework for efficient 3D depth profile estimation.
• To reduce computational complexity and memory requirements for real-time 3D reconstruction.
• To demonstrate accurate 3D profiling using synthetic training data.

Main Methods:

• Developed LiteF2DNet, a lightweight deep learning framework with dense connections for feature extraction.
• Employed synthetic data training using Gaussian mixture models and CAD objects to avoid real sample collection.
• Evaluated performance against standard methods using qualitative and quantitative analyses.

Main Results:

• LiteF2DNet achieved a 40% reduction in parameters compared to base models, enabling faster inference and lower memory usage.
• The model demonstrated superior performance in high dynamic ranges, with low-frequency fringes, and high noise conditions.
• Accurate 3D profiles of real objects were predicted using models trained solely on synthetic data.

Conclusions:

• LiteF2DNet offers an efficient and effective solution for real-time 3D reconstruction via fringe projection profilometry.
• Synthetic data training is a viable strategy for developing robust deep learning models for 3D profiling.
• The proposed lightweight framework is suitable for applications requiring limited computational resources.