Low-rank quaternion matrix representation has proven to be a powerful tool for color image restoration, excelling at preserving the inherent correlations among color channels. Conversely, while deep learning-based methods often deliver superior performance, they are hampered by critical drawbacks, including a dependence on large-scale datasets, high computational costs, and a lack of interpretability. To bridge the gap between these model-driven and data-driven paradigms, this talk introduces a novel framework that embeds a deep denoiser as a powerful prior within a low-rank quaternion optimization model. This hybrid approach synergistically combines the interpretability and robustness of traditional models with the expressive power of deep learning. Experimental results demonstrate that our proposed method achieves remarkable restoration quality by effectively leveraging the strengths of both worlds. The talk will conclude by outlining promising directions for future research in this domain.