An efficient CNN-based regression model for noise prediction in color images

Abstract

The presence of noise in digital images remains a critical barrier to high-fidelity visual representation, compromising both perceptual quality and the accuracy of downstream image analysis. Among various noise types, Gaussian and Poisson noise stemming from electronic sensor imperfections and photon-counting variability, respectively pose significant challenges. Accurate estimation of these noise levels is essential for improving computer vision tasks such as denoising, super-resolution, segmentation, and object detection. However, existing approaches often struggle under high-noise conditions, mixed-noise scenarios, and color image contexts. To address these limitations, we introduce two deep regression models: NoiseNet and NoiseNetV2, designed with noise-specific feature extraction and attention mechanisms to enhance performance across diverse noise types and datasets. Evaluated on datasets including Flickr30k, COCO, CelebA, and DIV2K, the proposed NoiseNet model achieved state-of-the-art results. For Gaussian noise, it recorded the Mean Absolute Error (MAE) of 0.0038, a Root Mean Squared Error (RMSE) of 0.0052, and an R² score of 0.9910, outperforming the best-performing baseline DenseNet121 by margins of ~2.7× lower MAE, ~2.5× lower RMSE, and ~1.05× higher R² score. For Poisson noise, NoiseNet achieved an MAE of 0.0547, RMSE of 0.0763, and R² of 0.9922, representing improvements of ~ 4× lower MAE, ~3.7× lower RMSE, and ~1.1× higher R² score compared to the best-performing baseline ResNet50. The model also outperformed classical techniques like BM3D and Scikit-learn in both low- and high-noise scenarios. Furthermore, NoiseNetV2, an enhanced variant incorporating an attention mechanism, delivered even stronger results under mixed noise conditions. For Gaussian noise, it reached an MAE of 0.0078, R² of 0.9682, and RMSE of 0.0103; for Poisson noise, it obtained an MAE of 0.2759, R² of 0.8106, and RMSE of 0.3763, outperforming both NoiseNet and all other baseline models for mixed noise. These findings confirm the proposed models’ strong generalization and practical utility in fields such as medical imaging, autonomous driving, remote sensing, and surveillance, enabling precise noise estimation for advanced image restoration and real-time vision systems.