IQI-UNet: A Robust Gaussian Noise Removal in Image Quality Improvement]{IQI-UNet: A Robust Gaussian Noise Removal in Image Quality Improvement
Sonda Ammar, Amina Kchaou, Bassem Ben Hamed · Research Square · 2024
Abstract Image quality enhancement is a rapidly advancing field in computer vision, with researchers exploring various techniques. This paper introduces a robust Gaussian noise removal, denoted as IQI-UNet, based on the U-Net deep learning architecture, which excels in image denoising. The key advantage of our model is its ability to operate without a reference image or prior knowledge of the application context. Our study investigates the impact of incorporating additional blocks into the autoencoder, resulting in an optimized architecture. The performance of the proposed IQI-UNet model is evaluated against twelve well-known methods using three datasets, namely Set5, Set14, and CBSD68. The proposed model demonstrates superior performance compared to established methods, achieving a remarkable 43% improvement in PSNR and a 29% improvement in SSIM. Furthermore, we introduce Shannon entropy as a metric to assess both the quantity and the quality of retained data. The evaluation, conducted on benchmark datasets, reaffirms the efficacy of IQI-UNet in enhancing image quality. The outcomes are highlighted by statistical tests, which implies that the obtained results are statistically significant.