Hello, my name is Simon Grosche and I'm glad to present to you our latest paper called

novel consistency check for fast recursive reconstruction of non-regular sampled video

data.

This is joint work together with Jürgen Seiler and Andrej Kaupp.

I would like to shortly go through all the highlights of our poster live presentation.

First, I'd like to introduce you to the core problem we are dealing with.

It is based on non-regular quarter sampling, which was first introduced in 2011.

Using non-regular sampling, here the spatial resolution or image quality per pixel can

be increased.

It can be understood as a special case of complex sensing as shown in one of our latest

contributions from 2020.

For video data, additional information can also be used from past frames.

This is highlighted in this sketch.

Here you can see three frames in each different pixel locations are measured.

You can think of a sensor which can measure different pixels in each frame.

Now the pixel positions from the past frames can also be projected to the current frame

in order to support the reconstruction.

This can be done using a motion estimation first with the non-regular sample data.

And of course, as those of you who are familiar with motion estimation know that it is useful

to use consistency checks to sort out low quality projections.

Now as a state of the art, such a problem is tackled using two algorithms.

One is the single frame frequency selective reconstruction, which is quite successful

for image data.

However, it does not make use of the projections and therefore was outperformed by the so-called

recursive FSR, so recursive frequency selective reconstruction in 2016 already.

Here one limitation is that only a fixed mass is used other than what I presented you

so far.

However, for this fixed mask, this exactly does what we explained before.

We do a motion estimation, then a projection of the pixels from the past and also they

applied several consistency checks before the projection.

You can see this in this graph.

So for the single frame FSR, only the data from the current measurements, from the currently

measured frames are used in order to reconstruct the image.

For the recursive FSR, but also for the later explained DFSR, we use data from all three

past frames here in this illustration in order to reconstruct the current image.

Now let me go through our contributions.

Our novel contributions are summarized in the so-called DFSR, which stands for dynamic

FSR.

This is a generalization to dynamically changing sampling mass.

So other than the recursive FSR, which can only deal with a fixed sampling pattern, we

can change the sampling pattern every frame.

The DFSR is faster because it uses faster consistency checks.

And this also, and these consistency checks are not only faster, but they also result

in higher quality.

I will not go into the details of these three new consistency checks.

Just to summarize, this is the fastest version, this is just a fast version of one of our

consistency checks.

And also this NNC consistency check is quite fast.

And we will see that the combination of these two gives us high quality and speed.

So with this, I already like to go through the simulations and results.