<p>Artificial intelligence and deep learning techniques are increasingly transforming image enhancement and quality assessment by improving perceptual quality and structural fidelity across diverse imaging applications. However, many existing approaches rely exclusively on either full-reference or no-reference image quality metrics, which limits their effectiveness in real-world scenarios where reference images may not always be available. To address this limitation, this paper proposes a deep neural network (DNN) framework that integrates multi-scale convolutional feature extraction with a hybrid perceptual optimization strategy combining full-reference and no-reference quality metrics. The proposed architecture consists of hierarchical convolutional layers for multi-scale feature representation, followed by a perceptual quality optimization module that incorporates structural similarity constraints, fidelity-based loss functions, and learned no-reference quality predictors to guide image enhancement. To demonstrate the robustness and generalization capability of the proposed framework, experiments are conducted on both medical imaging datasets (NIH Chest X-ray14, CheXpert, and Brain Tumor MRI datasets) and natural image benchmarks including DIV2K and BSD500, thereby validating the model across heterogeneous visual domains. Experimental results show that the proposed approach consistently outperforms existing methods in terms of objective image quality metrics such as PSNR, SSIM, and LPIPS, while also achieving strong classification performance on medical datasets. In addition to high accuracy, the models improved sensitivity and specificity while maintaining computational efficiency, making them suitable for practical image quality assessment and enhancement tasks. These results demonstrate that integrating reference-dependent and reference-independent quality cues within a unified deep learning framework significantly improves perceptual image quality and robustness across diverse imaging domains.</p>