A Gradient-Projected Model for Image Denoising.

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Digital images are susceptible to noise during acquisition, degrading structural information and hindering analysis.
• Existing denoising methods often struggle to balance noise reduction with the preservation of fine image features.

Purpose of the Study:

• To introduce AuroraNet, an advanced image denoising framework designed for robust performance on real-world noisy images.
• To enhance training stability and preserve fine-scale image features through a novel optimization technique.

Main Methods:

• AuroraNet extends the dual-branch architecture of DudeNet.
• Integration of a Gradient-projected Function (GPF) optimizer for improved training dynamics.
• Evaluation on two real-world noisy image datasets under diverse noise conditions.

Main Results:

• AuroraNet achieved high Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) scores (e.g., 38.40 dB PSNR, 0.9633 SSIM).
• Consistently outperformed established denoising models and the baseline DudeNet in reconstruction quality.
• Demonstrated comparable performance to R-REDNet with significantly fewer parameters, highlighting computational efficiency.

Conclusions:

• AuroraNet presents a computationally efficient and effective solution for real-world image denoising.
• The framework successfully balances strong denoising capabilities with a reduced parameter count.
• Offers practical value for applications requiring high-quality image reconstruction without excessive computational cost.