Consumer-grade depth sensors often generate low-quality, low-resolution depth images. Leveraging the correlation between depth and high-resolution RGB images presents a promising solution. While current methods struggle to capture the complex and dynamic relationship between these modalities, we propose a novel weighted analysis representation model for enhanced depth image processing. Our approach incorporates task-driven learning and dynamic guidance. By introducing a guided weight function, we refine the analysis representation model to better capture dependencies between depth and RGB images. Task-specific optimization is achieved through a task-driven learning framework. Moreover, to adapt to the evolving depth image quality, we employ dynamic guidance, where stage-wise parameters are learned to adjust guidance signals iteratively. The efficacy of our method is demonstrated through applications in depth image upsampling and noise reduction.