Image restoration is a fundamental task in image processing with wide-ranging applications in modern life, including medical imaging, remote sensing, radar imaging, and digital preservation of historical and museum artifacts. The objective of image restoration is to recover a high-quality image from degraded observation by reducing the effects of noise and blur. Effective restoration depends on understanding the degradation process; therefore, identifying the type of noise and the blur model is essential. In practical scenarios, images are often degraded by atmospheric and environmental conditions and restoring them requires appropriate restoration techniques tailored to the distortion characteristics. This paper reviews and assesses contemporary machine learning-based image restoration methods. The proposed evaluation reports quantitative performance across four standard benchmark datasets Kodak24, CBSD68, Urban100, and LIVE—using PSNR (dB), MSE, and SSIM as primary quality metrics. The achieved PSNR scores are 27.24 dB, 29.38 dB, 30.04 dB, and 30.91 dB on Kodak24, CBSD68, Urban100, and LIVE, respectively. The corresponding MSE values are 367.56, 224.88, 193.10, and 158.02, while SSIM values are 0.8690, 0.9337, 0.9432, and 0.8008. These results demonstrate the effectiveness of the evaluated approach in improving image quality across diverse image restoration benchmarks.
Kumavat et al. (Thu,) studied this question.