This talk primarily explores how to embed deep learning priors into traditional variational optimization model frameworks to effectively enhance the accuracy, generalizability, and interpretability of deep learning methods. The talk mainly covers two aspects: 1) Proposing a Deep Embedded Variational Optimization (DE-VO) framework which is a general form to connect traditional and DL methods; 2) Showing some specific examples under the DE-VO framework. The proposed framework exhibits strong interpretability and demonstrates superior accuracy and generalizability in high-dimensional data processing and analysis tasks such as remote sensing image fusion.