Robust Focus Volume Regularization in Shape From Focus.

Area of Science:

• Computer Vision
• Image Processing
• 3D Reconstruction

Background:

• Shape from focus (SFF) reconstructs 3D scenes from multi-focus images, with quality dependent on focus volume (FV) accuracy.
• Traditional SFF methods struggle with edge preservation, fine detail reconstruction, and artifact removal, often lacking shape priors.

Purpose of the Study:

• To enhance the accuracy of 3D shape reconstruction by refining the focus volume (FV) in Shape from Focus (SFF) techniques.
• To address limitations of traditional SFF methods by incorporating a nonconvex regularizer and dual shape priors for improved performance.

Main Methods:

• Developed an energy minimization framework utilizing a nonconvex regularizer robust to noisy focus values.
• Incorporated two types of shape priors: a static prior from the input image sequence and iterative, on-the-fly FVs.
• Optimized the nonconvex energy function using the majorize-minimization algorithm for guaranteed local minima and rapid convergence.

Main Results:

• The proposed method demonstrates superior accuracy in reconstructing 3D shapes compared to existing approaches.
• Experimental results on synthetic and real image sequences validate the method's effectiveness in preserving structural edges and fine details.
• The approach effectively removes noisy artifacts while maintaining high-fidelity 3D shape representation.

Conclusions:

• The novel SFF approach, leveraging nonconvex regularization and dual shape priors, significantly improves 3D shape reconstruction quality.
• The method offers a robust and accurate solution for detailed 3D shape recovery from multi-focus image sequences.
• This work advances SFF techniques by providing a more effective way to handle noisy data and incorporate prior shape information.