<p>Single image super-resolution (SISR) is a crucial aspect of computer vision and image processing, aiming to restore high sharpness and visual quality to low-resolution images. It has practical applications in facial recognition, video surveillance, medical imaging, remote sensing, and scene text recognition. SR techniques, including conventional and machine learning-based algorithms, have been presented in the literature. Conventional methods often fail to recreate fine details, while learning-based techniques, particularly those using residual networks, show impressive results in image-enhancing tasks. Nevertheless, larger unnecessary architecture tends to bring about the growing weight scale of model, execution time and memory consumption etc. To resolve the problems mentioned above, we propose to construct new mean-std normalization-based networks named as Multi-Scale Deep Residual Networks (MSDRNs) and develop a Bayesian hyperparameter optimization approach, aiming to improve residual learning with respect to stability and generalization. The Mean-Std Normalization enforces uniform statistical scaling of all residual branches and the Bayesian optimization adaptively tunes critical parameters to achieve robust convergence. The research combines full-reference metrics (PSNR, SSIM) with no-reference NIQE index to evaluate overall quality of reconstructed images. It combines quantitative accuracy with human-visual fidelity. When comparing with linear, residual, recursive, dense and attention-based approaches it is observed that the proposed solution for an image divination (SSIM = 0.94), naturalness preservation (NIQE = 6.73) significantly better in terms of quality (better than SwinIR and EDSR). The proposed method exploits the high resolution for accuracy maximization.</p>