Document Type : Research
Author
Department of Physics, Bonab Center, Islamic Azad University, Bonab, Iran.
Abstract
Poisson-Gaussian mixed noise removal in optical sensor images is considered one of the major challenges in digital imaging, machine vision, and photon-limited optical systems due to its heterogeneous statistical nature and the signal-dependent characteristics of part of the noise. In this study, a deep convolutional neural network (DnCNN) based on a residual learning strategy was developed for image restoration and denoising. The proposed architecture, incorporating batch normalization layers and the Adam optimization algorithm, exhibited favorable stability and convergence during the training process.
The simulation results and both quantitative and qualitative evaluations demonstrated that the proposed network outperformed conventional denoising methods in suppressing mixed Poisson-Gaussian noise and restoring image information. In addition to improving image quality assessment metrics, the proposed method showed a high capability for preserving high-frequency structures, edge details, and photometric features. Furthermore, the performance analysis of the model under different noise levels indicated its stability, satisfactory generalization capability, and robust performance under low-light and high-noise imaging conditions. The findings of this study indicate that the DnCNN-based approach effectively removes mixed noise while preventing image blurring and loss of sharpness, thereby enabling more accurate recovery of photometric and structural information. Therefore, the proposed approach can serve as an efficient computational framework for low-light imaging systems and for scientific and industrial applications based on optical sensors.
Keywords
)