Review Study on Inverse Problems Using Deep Learning Techniques with applications

Current research in machine learning indicates that deep neural networks have the ability to address complex challenges in computational imaging, specifically referred to as inverse problems. The inverse problems in image reconstruction and image recovery have the significant importance in the tomographic imaging applications. In this review paper, we introduce a comparison between different regularization techniques employed in the literature for image recovery. Furthermore, this involves comparing various numerical optimization algorithms and techniques used in image reconstruction processes, which includes integrating the conjugate gradients algorithm within the network. These techniques encompass Conjugate Gradient-based Splitting Total Variation (CSTV), Splitting Douglas-Rachford with Edge-Thresholding and Weighted Shrinkage (SD-ET-WS), and Conjugate Gradient - Primal Dual - Nesterov Scheme (CG-PD-NS). These models depend on proximal gradient steps to maintain data consistency. Furthermore, a comparison between different deep learning frameworks is introduced to investigate the best model performance in terms of acceleration factors.