This paper proposes a deep CNN denoiser prior for image restoration, combining model-based optimization with discriminative learning. The authors train a set of fast and effective CNN denoisers and integrate them into model-based optimization methods to solve various inverse problems, including image deblurring and super-resolution. The CNN denoisers are trained using techniques such as ReLU, batch normalization, and dilated convolution, enabling efficient and effective denoising. The denoiser prior is integrated into the optimization framework, allowing the use of fast CNN-based methods for image restoration tasks. Experimental results show that the proposed method achieves promising Gaussian denoising results and performs well in various low-level vision applications. The method also demonstrates competitive performance in image deblurring and super-resolution tasks, outperforming existing methods in terms of PSNR and computational efficiency. The integration of model-based optimization and discriminative learning methods offers a flexible and effective framework for image restoration. The study highlights the potential benefits of combining these two approaches, demonstrating that learning expressive CNN denoiser priors can be a good alternative to traditional image priors. The work also provides insights into designing CNN architectures for task-specific discriminative learning.This paper proposes a deep CNN denoiser prior for image restoration, combining model-based optimization with discriminative learning. The authors train a set of fast and effective CNN denoisers and integrate them into model-based optimization methods to solve various inverse problems, including image deblurring and super-resolution. The CNN denoisers are trained using techniques such as ReLU, batch normalization, and dilated convolution, enabling efficient and effective denoising. The denoiser prior is integrated into the optimization framework, allowing the use of fast CNN-based methods for image restoration tasks. Experimental results show that the proposed method achieves promising Gaussian denoising results and performs well in various low-level vision applications. The method also demonstrates competitive performance in image deblurring and super-resolution tasks, outperforming existing methods in terms of PSNR and computational efficiency. The integration of model-based optimization and discriminative learning methods offers a flexible and effective framework for image restoration. The study highlights the potential benefits of combining these two approaches, demonstrating that learning expressive CNN denoiser priors can be a good alternative to traditional image priors. The work also provides insights into designing CNN architectures for task-specific discriminative learning.