Abstract
Digital images are prone to various forms of noise during acquisition, which can distort structural information and hinder subsequent processing. This work proposes AuroraNet, a denoising framework that extends the dual-branch design of DudeNet and integrates a Gradient-projected Function (GPF) optimizer to enhance training stability and preserve fine-scale image features. We evaluated the model on two real-world noisy image datasets to examine its robustness under different noise conditions. AuroraNet achieved an average PSNR of 35.59 dB on the first dataset and 38.40 dB on the second, together with an SSIM of 0.9633 in the latter. Across both benchmarks, AuroraNet consistently delivered higher reconstruction quality than several established models and the baseline DudeNet. Although R-REDNet produced the highest overall scores on one of the datasets, AuroraNet attained comparable performance while using a much smaller amount of parameters, underscoring its efficiency and practical value. These results indicate that AuroraNet offers a balanced solution for real-world image denoising, providing strong denoising capability without sacrificing computational economy.