Low-Rank Matrix Completion to Reconstruct Incomplete Rendering Images.

Area of Science:

• Computer Graphics
• Image Processing
• Computational Imaging

Background:

• Path tracing offers realistic rendering but suffers from noise in previsualization due to insufficient samples.
• Reconstructing missing samples is crucial for improving the quality of intermediate rendering outputs.

Purpose of the Study:

• To develop novel methods for reconstructing missing samples in incomplete images for path tracing.
• To enhance the visual quality of previsualization in computer graphics applications.

Main Methods:

• Utilizing a low-rank matrix completion framework to exploit visual signal coherence.
• Employing a convolutional neural network for fast pre-completion of missing pixel values.
• Applying weighted nuclear norm minimization (WNNM) with a parameter adjustment strategy (PAWNNM) for efficient reconstruction.

Main Results:

• Successfully reconstructed missing pixels, sub-pixels, and multi-frame scenarios in images.
• Achieved superior visual quality compared to recent methods, including compressed sensing based reconstruction.
• Demonstrated efficient recovery of missing values, preserving high-frequency details.

Conclusions:

• The proposed low-rank matrix completion framework effectively addresses noise in path tracing previsualization.
• The combination of CNN pre-completion and PAWNNM offers a robust solution for image reconstruction.
• This approach significantly enhances the quality and usability of intermediate rendering results.