<p>The advancement of convolutional neural networks (CNNs) has significantly improved single-image super-resolution (SISR). While most existing methods yield promising results under Bicubic down-sampling degradation, they struggle in realistic scenarios where low-resolution (LR) images are corrupted by multiple degradations, including additive white Gaussian noise (AWGN). Discriminative learning-based models produce high-quality outputs but are highly dependent on training data, making them vulnerable to complex conditions. In contrast, model-based approaches such as maximum a posteriori (MAP) provide greater flexibility but rely on handcrafted priors, which may be inadequate. To address these limitations, we propose a hybrid non-blind SR framework based on half-quadratic splitting (HQS), embedding CNN-based learning into both prior and likelihood terms within an iterative model. Unlike other hybrid methods that consider learning only in the prior term, our approach leverages the strengths of CNNs and MAP by incorporating learning capabilities in both terms. The prior model operates in both spatial and frequency domains to extract local and global features using enhanced ghost blocks with ghost attention (GBGA), which reduce feature redundancy and computational cost while emphasizing informative regions. Additionally, a CNN with residual connections across multiple convolutional layers enables learning for the data term in a shallow-to-deep information flow structure. Experimental results demonstrate the effectiveness of the proposed model in both qualitative and quantitative evaluations. It achieves competitive performance under Bicubic degradation, excels in simultaneous super-resolution and denoising—achieving up to 0.074 dB improvement over state-of-the-art methods—and remains robust under noise levels exceeding those encountered during training.</p>