Farhad Niknam*
Physics-based image formation models enable computationally obtaining meaningful information by processing other forms of information which can be acquired through measurements. In practical situations however, the inner functionalities of the system which create the impulse response function are usually unknown, and due to noise, measurements are unreliable. Before Deep Neural Networks (DNNs) taking over, Compressed Sensing (CS) techniques were primarily being used to address this lack of information by imposing assumptions into the problem. But this switch to DNNs came with the price of mass data acquisition for training to leap over the never-ending problem of algorithmic fidelity in CS methods. Recently, deep image prior and untrained or semi-trained networks, while leveraging the power of DNNs and algorithms, have become successful to be considered as potential answers to the desire of finding a cost-efficient yet powerful solution. In this paper, we briefly have a look at the recent breakthroughs conducted over this concept to solve various imaging problems.
分享此文章
English 
Spanish 
Russian 
German 
French 
Japanese 
Portuguese 
Hindi 