Deep Image Prior for solving inverse problems in imaging methods
Deep Image Prior (DIP) introduced a new way of using Artificial Neural Networks that dismisses the need for training its parameters on a large dataset. This method allows solving inverse problems in imaging using only the measurements of the case to be solved. This thesis aimed to investigate whether DIP is a worthy method to explore for solving modern inverse problems in imaging. We tested the method for reconstructing Electrical Impedance Tomography images of the human head, restoring severely blurred text images, and reconstructing limited-angle Computed Tomography images. We also proposed modifications to the technique to improve its performance. Our results showed that DIP achieves good performance even for moderately high levels of data degradation, surpassing traditional methods and some implementations of supervised Deep Learning algorithms. We were also able to improve upon the method's original proposal with our modifications. However, we could not achieve the performance of end-to-end Deep Learning approaches. In conclusion, we found that DIP is a powerful and versatile method and is a good candidate to be explored for solving inverse problems in imaging, especially when gathering large datasets is not feasible.