Self-Supervised Contrastive Learning and GAN-Based Denoising for High-Fidelity HumanNeRF Images

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• HumanNeRF generates realistic 3D human models but suffers from image noise and detail loss.
• This degradation stems from incomplete training data and rendering process sampling noise.

Purpose of the Study:

• To develop an effective image denoising method for HumanNeRF-generated images.
• To enhance detail fidelity and overall image realism.

Main Methods:

• Utilized self-supervised contrastive learning to differentiate noise from human details without external labels.
• Employed Generative Adversarial Networks (GANs) for adversarial training to refine image realism and detail representation.

Main Results:

• Successfully removed noise from HumanNeRF images.
• Significantly improved detail fidelity and image quality.
• Demonstrated superior performance in enhancing human image realism.

Conclusions:

• The proposed method effectively denoises HumanNeRF images.
• It enhances detail fidelity, supporting improved 3D human reconstruction and rendering.
• Combines self-supervised learning and GANs for robust image enhancement.