Computer Vision and Artificial Intelligence Lab
Deep learning provides the best performing solutions for many problems in computer vision and image processing. There are several work in literature that show the superiority of deep learning over classical variational methods in inverse problems, such as image denoising, sharpening and super-resolution. In this project, we will focus on the solution of various inverse problems using a general deep learning framework which does not depend on the physical model of the problem. The most important novel contribution of the project is that the developed deep architecture will be almost independent of model parameters and will be easily adaptable to different problems, which differentiates it from methods in literature. For that purpose, separate deep architectures will be implemented for model estimation, reconstruction and regularization steps of classical iterative optimization techniques. Then these architectures will be combined and the whole system will be trained end-to-end. This project will investigate and provide solutions for novel problems that deal with the structure, training and interaction of these architectures.
Publications:
- Deep Learning-Based Blind Image Super-Resolution using Iterative Networks, Visual Communications and Image Processing (VCIP 2021).
Visual Results for BISR-Net
(compared with DAN: https://github.com/greatlog/DAN )
BISR-Net 
DAN
BISR-Net 
DAN