Document Type : Research

Author

• Ali Raheli

Department of Physics, Bonab Center, Islamic Azad University, Bonab, Iran.

Abstract

Removing mixed Poisson–Gaussian noise is a challenging task in optical sensor image processing because the variance of its signal-dependent component varies with image intensity. In this study, a residual-learning convolutional neural network comprising 11 convolutional layers and batch normalization was developed to estimate noise and reconstruct images, and was trained using the Adam optimizer. Performance was evaluated on 24 synthetic images held out from training under nine noise conditions, with three noise realizations per image, and compared with Gaussian and median filtering. Both training and validation losses decreased over 20 epochs. At (p=100) and (σ=0.03), the network achieved a mean peak signal-to-noise ratio (PSNR) of 31.12 dB and a structural similarity index (SSIM) of 0.875. The PSNR improvements over the noisy input and median filtering were 5.93 and 0.72 dB, respectively. However, Gaussian filtering achieved a comparable PSNR of 31.08 dB and a higher SSIM of 0.887. Across the nine noise conditions, the network achieved a higher mean PSNR than median filtering in all cases and outperformed Gaussian filtering in five cases. Visual assessment also indicated reduced noise and partial preservation of the main image structures, although localized reconstruction errors remained.

Keywords

• Optical sensors
• Poisson-Gaussian mixed noise
• Deep learning
• Residual learning
• Image processing