MINet: A DENOISING APPROACH FOR MEDICAL AND SATELLITE IMAGING APPLICATIONS

Image denoising is a crucial step in various applications, aiming to remove noise while maintaining the important features. Unwanted noise is typically present in medical and remote sensing satellite images, which compromises the accuracy of the diagnosis, and it is extremely difficult to design an efficient network for image denoising with better performance and fewer parameters. However, the conventional techniques employed for denoising were prone to specific problems, namely computational complexity, data loss from compression, and edge detail preservation. Hence, this research proposed a multilevel imperialistic U-Net (MINet) model that integrates the advantages of nature-inspired optimization and deep learning techniques for medical and remote sensing image denoising. The proposed imperialistic learning optimization (ILO) algorithm is the combination of teaching-learning and imperialistic behaviors, which excels at local search, finding potentially good solutions to tune the hyperparameters, and ensuring robust performance across various noise levels. Leveraging the strength of two behaviors solely maximizes convergence speed and minimizes overfitting. On top of that, the MINet’s resilient architecture enables meticulous denoising with minimum loss of structural information, gamma correction modifies the illumination for improved visibility, and two-level diffusion assists in smoothing the noisy areas, which in turn aids in enhancing the overall image quality. As a result, the MINet model achieves a structural similarity index (SSIM) of random noise of 0.77 for medical images and 0.74 for satellite images, outperforming other conventional methods. The proposed approach effectively removes noise while maintaining fine details.