Depth map filtering method for shape-from-focus recovery based on hybrid network model

Area of Science:

• Optics and Photonics
• Computer Vision and Image Processing
• Metrology

Background:

• Shape-from-focus (SFF) is vital for 3D shape recovery, particularly in digital microscopy.
• Image texture richness, essential for SFF, often introduces noise into depth maps, degrading quality.
• Existing methods struggle to effectively filter noise without compromising structural integrity.

Purpose of the Study:

• To develop an advanced depth map filtering framework for shape-from-focus (SFF) systems.
• To address the critical challenge of noise reduction in SFF-generated depth maps.
• To improve the accuracy and reliability of 3D shape recovery in textured environments.

Main Methods:

• A hybrid network model was designed for focus signal classification and noise pattern identification.
• An initial depth map was generated using a combination of Full Width at Half Maximum (FWHM) and Gaussian fitting.
• A flag matrix was employed to guide the filtering process, selectively removing noise.

Main Results:

• The proposed method effectively distinguishes focus signal patterns, identifying noise introduced by signal distortion.
• Experimental results show significant noise reduction in depth maps.
• Structural details within the depth maps were preserved, demonstrating excellent filtering performance.

Conclusions:

• The developed framework provides a robust solution for filtering noisy depth maps in SFF applications.
• The method successfully filters noise while maintaining the integrity of 3D structural information.
• This approach offers a general SFF framework applicable to diverse 3D shape measurement and instrument design fields.