Understanding the Deep Image Prior

In 2018 a paper called “Deep Image Prior” was published . The paper presented a method of using a convolutional network as a prior for various image restoration tasks. In this method, the training of the convolutional network was done using a single image (the image to be restored), without need of a big data set. This new novel technique gained both positive academic attention and criticism for not addressing fundamental questions and supposably “cherry picking” results.
Our project had two main goals:
The first was to deepen our understanding of the work presented in the paper. By diving into its implementation, we tried to answer questions that the original paper did not refer to. We have found different code parts that were not mentioned in the paper, and, once removed, vastly lowered the quality of the output. Furthermore, we researched unwanted phenomena that were not mentioned in the paper.
Secondly, we strive to generalize and improve the paper's work, partially by checked the effect of different parameters and architectures on the results. Lastly, we suggested a solution to the main issues that leads to better results than the original paper.

Understanding the Deep Image Prior
Understanding the Deep Image Prior
Understanding the Deep Image Prior