Search

Search Results

• NOISE GENERATION METHOD BASED ON A SET OF NOISY IMAGES WITHOUT CLEAN EXAMPLES

  А.S. Kovalenko , Y. М. Demyanenko

  243-254

  2025-11-10

  Abstract ▼

  In this work, a novel method is proposed for noise generation from noisy images that does not require aligned pairs of clean and noisy data. Unlike traditional approaches demanding matched image sets or a priori noise models, the developed technique models complex noise characteristics intrinsic to specific CMOS sensors solely from observed noisy data. Noise synthesis is achieved via a U‑Net‑like generative adversarial architecture based on StyleGANv2, featuring a modified discriminator conditioned on camera parameters and input image metadata. Special emphasis is placed on preserving the spatial–color structure and textural details of each image, enforced through a dedicated loss function that ensures fidelity to the original color rendering and fine-grained patterns. Training of the noise generator is performed without any paired clean and noisy images, which proves particularly valuable when handling real-world datasets acquired from multiple camera models under varied lighting conditions. The experimental section presents a detailed comparative analysis of the synthesized images using PSNR and SSIM metrics, along with an evaluation of the noise distribution based on intensity statistics and spectral characteristics. It is demonstrated that the generated dataset functions effectively as a standalone training corpus for denoising neural networks and, when combined with a real dataset (e.g., SIDD), yields further enhancements in denoising performance. Results indicate that combined training on the union of generated and real examples produces an average PSNR improvement of 1.5 dB compared to existing methods reliant on aligned data. Independence from the specific optical characteristics of any given sensor significantly broadens the method’s applicability. These findings confirm the utility of the proposed approach for realistic noise synthesis and removal in scenarios lacking clean reference images, and they open avenues for future research into adaptive noise-model generation