So, welcome everybody.

Let's continue with our lecture on interventional medical image processing.

And today we will look into some noise reduction techniques and the interesting part is that

these noise reduction techniques will try to preserve edges because we've already seen

edges are important and noise kind of interferes with the processes that we want to apply.

For example, if we want to detect key points, we've already seen that edge preserving filtering

is a good post-processing technique in order to improve the outcomes of the key point detection.

So today we will look into several versions of edge preserving filtering and we will have

two methods in particular which we will highlight.

One is based on bilateral filters and the other one is based on guided filtering.

Both techniques are very popular and they are quite straightforward to implement and

we will also show towards the end of the lecture how we can adapt those methods to a particular

application.

And you will see that most of the time it's not so difficult to adapt the algorithm but

you still have to work with the algorithm and you have to understand the properties

of your data and then you can get pretty nice processing results.

So here the motivation is going to be range imaging, later we will also look into x-ray

imaging for the edge preserving filtering.

Of course you've already seen this kind of sensors here in the lecture, we've been working

with them quite a bit from range imaging data.

You can typically get different kinds of information.

The nice thing is for example with time of flight camera, so a time of flight camera

actively measures at every pixel how long the time took for a light source to emit a

particular light ray and the time until it's measured back.

So this essentially is a depth measurement because you can compute from the time that

the light propagates and returns to the sensor you can compute a depth value at every pixel

or you can use structured light like in the old Kinect sensor that has been employed using

some kind of random pattern that is projected into the scene and then you can use point

correspondences to reconstruct the depth at every point of the random pattern and from

that you can also get a dense depth field.

So the nice thing is those camera technologies deliver not just an RGB image or a grayscale

image but they also deliver the depth at every pixel in the camera image.

So you have a dense field that can display often this kind of information is called 2.5D

because you get essentially 3D information but only from your viewpoint.

So they call it 2.5D or because the depth dimension is so you don't have like in a 3D

data set if you have a slice image you have full 3D information but here you only get

the information until the surface until you have the first interaction of the scene.

And on the right side you see a couple of examples so this is such a depth image and

here the depth is color coded and you can see that this hand is in front of this hand.

And you also get some grayscale information so this is the grayscale image where you can

get some texture features from and this is typically aligned to those two information

sources or you can calibrate them against each other.

And the other thing is of course from the range data you can reconstruct a point cloud

and this point cloud shows you the depth and if you have a visualization in 3D then you

can even turn this image and you can create a lateral view onto your point cloud and of

course you only get data on the surface with respect to the viewing direction.

Typically those cameras have a measurement range in between let's say one to seven meters

or one to three meters so this is the typical range that the depth data is acquired and

you have only rather low resolutions like 200 by 200 pixels or 640 by 480 so this is a typical

configuration for such a camera.