Real-Time Volume-Rendering Image Denoising Based on Spatiotemporal Weighted Kernel Prediction.

Area of Science:

• Computer Graphics
• Image Denoising
• Machine Learning

Background:

• Volumetric Path Tracing (VPT) using Monte Carlo (MC) sampling generates noisy images, especially in real-time applications due to limited samples per pixel.
• Existing real-time denoising methods struggle with temporal stability and detail preservation, leading to blurry results.

Purpose of the Study:

• To develop a lightweight, spatiotemporal neural network for effective denoising of low-sample VPT images.
• To enhance image quality and temporal stability in real-time rendering scenarios.

Main Methods:

• Utilized a reprojection technique to extract features from historical frames.
• Designed a dual-input convolutional neural network (CNN) to predict filtering kernels by independently encoding radiance and geometric features.
• Applied learned weight filtering kernels for spatiotemporal filtering of images.

Main Results:

• The proposed network demonstrated superior noise suppression compared to baseline models.
• Achieved enhanced feature extraction and detail representation capabilities.
• Showcased improved temporal stability in denoised images.

Conclusions:

• The spatiotemporal lightweight neural network effectively denoises low-sample VPT images.
• The method offers a significant improvement in image quality and temporal stability for real-time graphics.
• Outperforms existing denoising techniques in detail preservation and noise reduction.