Abstract Image denoising and sharpening are fundamental tasks in many computer vision applications. The current paper presents an unprecedented deep learning strategy that combines both denoising and sharpening functions in the same framework, achieving remarkable image quality measures. First, a denoising model based on the U-Net was presented with better performance (measured in terms of PSNR and SSIM) in comparison with classic state-of-the-art approaches such as BM3D, as well as recent deep learning methods including DnCNN and SwinIR. As the next step, the second model was specifically trained with the sharpening function, achieving an increment in the PSNR and SSIM measures when they were applied in the denoising model output. The main contribution of this paper is the introduction of the single end-to-end deep learning model that conducts both the denoising and sharpening functions in an automated way. The combined framework that benefits from the compact U-Net’s advantages in both denoising and the Restormer-like building block in sharpening achieves substantial performance improvements: for Gaussian noise, our unified model (UDNBS) attains an average PSNR of 25.09 dB and SSIM of 0.69, representing 15.7% and 50.0% improvements over BM3D (21.68 dB, 0.46 SSIM), 14.4% and 46.8% improvements over DnCNN (21.94 dB, 0.47 SSIM), and 10.0% and 40.8% improvements over SwinIR (22.82 dB, 0.49 SSIM), respectively. For speckle noise, UDNBS achieves 25.32 dB PSNR and 0.78 SSIM, demonstrating 4.2% PSNR and 6.8% SSIM gains over BM3D (24.30 dB, 0.73 SSIM), 3.0% PSNR and 1.3% SSIM gains over DnCNN (24.59 dB, 0.77 SSIM), and 0.6% PSNR gains with equivalent SSIM (0.78) compared to SwinIR (25.17 dB, 0.78 SSIM) across all noise variance levels tested. These quantitative improvements, coupled with 2.0–15.7% gains over sequential pipelines, are validated across multiple noise types (Gaussian, speckle, salt-and-pepper, and Poisson). Comprehensive experiments demonstrate that UDNBS consistently outperforms both traditional methods and recent deep learning approaches while maintaining computational efficiency through its compact architecture. The current paper proves the effectiveness of the merged solution in powerful image restoration with the introduction of a more efficient and superior solution for simultaneous resolution of issues related to the presence of noise and blur.
Cevik et al. (Wed,) studied this question.