Deep Learning-Based Blind Quality Assessment for Degraded Underwater Images

The evaluation of underwater image quality remains a challenging task due to the complex and unpredictable distortions introduced by light absorption, scattering, and color attenuation. Traditional full-reference and reduced-reference metrics are often impractical in real-world underwater scenarios, where pristine reference images are unavailable. This study proposes a deep learning-based blind quality assessment framework that automatically predicts the perceptual quality of degraded underwater images without requiring reference data. The proposed approach leverages convolutional neural networks to extract hierarchical features that capture both low-level degradations (e.g., blur, noise, color cast) and high-level perceptual attributes. To enhance robustness, the model is trained on a diverse dataset of underwater images with varying degradation levels and is optimized using a combination of regression loss and perceptual consistency constraints. Experimental results demonstrate that the proposed method achieves strong correlation with human subjective evaluations and outperforms conventional no-reference quality assessment techniques. The framework provides an efficient and scalable solution for real-time underwater image analysis, with potential applications in marine exploration, autonomous underwater vehicles, and environmental monitoring.