1: \begin{abstract}

2: Recent advances using deep neural networks (DNNs) for solving inverse problems in image processing have

3: significantly outperformed conventional optimization algorithm based methods.

4: Most works train DNNs to learn 

5: 1) forward models and image priors implicitly for direct mappings from given measurements to solutions,

6: 2) data-driven priors as proximal operators in conventional iterative algorithms, or

7: 3) forward models, priors and/or static stepsizes in unfolded structures of optimization iterations.

8: Here we investigate another way of utilizing convolutional neural network (CNN) for empirically 

9: accelerating conventional optimization for solving inverse problems in image processing.

10: We propose a CNN to yield parameters in optimization algorithms 

11: that have been chosen heuristically, 

12: but have shown to be crucial for good empirical performance.

13: Our CNN-incorporated scaled gradient projection methods, without compromising theoretical properties, 

14: significantly improve empirical convergence rate over 

15: conventional optimization based methods 

16: in large-scale inverse problems such as image inpainting, 

17: compressive image recovery with partial Fourier samples, 

18: deblurring and sparse view CT. 

19: During testing, our proposed methods dynamically select parameters every iterations 

20: to speed up convergence robustly for different degradation levels, noise, or regularization parameters 

21: as compared to direct mapping approach.

22: \end{abstract}

23: