Deep convolutional networks have become a popular tool for image generation and restoration. Generally, their excellent performance is imputed to their ability to learn realistic image priors from a large number of example images.
On the contrary, the recent paper shows that the structure of generator CNN itself is sufficient to capture a great deal of low-level image statistics.
On the seminar, we will describe how a randomly-initialized neural network can be used as a handcrafted prior with excellent results in standard inverse problems such as denoising, super-resolution, and inpainting.
The whole thing indeed works like magic and proves the importance of developing a good networks architectures, which go against the common narrative that explain the success of deep learning in image restoration.
Speakers: Stanislav Belyaev.
Presentation language: Еnglish.
Date and time: October 9th, 20:30-22:00.
Location: Times, white boards (4th floor).