Okay, the last topic of this chapter is the generation of the label image.

The question is essentially how do we represent the results of segmentation, how do we represent our homogeneous regions? Well, if we apply some kind of semantic segmentation, what we get is a binary image of foreground and background pixels that would look like this.

And we can represent such regions as a label image. So the idea is that we define a new digital image, which has the same size as the original image, but the grey values indicate to which region the pixel belongs.

And then, of course, if we have a segmentation of the foreground and background pixels, then the background region would be indicated by a value of zero. So all pixels containing a value of zero would still be considered to be background pixels.

And other than that, well, for every segment, we would have a grey value corresponding to the segment number that indicates it belongs to it. Typically, we would also store the first coordinate to have a faster access to all of the pixels corresponding to an object.

So what we would store are the positions of the first point we encounter for a region. If we scan the image row by row, this would be the first one for pixel one. And other than that, in the label image, well, we would also have the label image where the pixel value indicates the index of the region.

And it's generated like this. So we add the first and the last columns and rows to the background. And now for every pixel, having a certain grey value, we check its neighbors.

And that's always the next neighbor in the row and the next neighbor in the same column. And well, what we will do is we will scan the entire image in this way and we will have a look at what happens with foreground pixels.

So this pixel is having a grey value of 255, like in which. And well, the first case could be, well, if we checked the previous pixel in the same column and the previous pixel in the same row.

Now here we have the first pixel, or here we have a pixel which has a grey value of one, or 255, which blocks the foreground. And now we have a case where both neighboring pixels, N and W, do not correspond to a region.

They are background pixels. So in this case, we would initialize a new region here. Now in this case, we would consider that the pixel has a neighbor that does already belong to one region, the green region.

And so we assign it to the same region, even though the region up here does not correspond to that region here. Well, in this case, we see that the point A actually has a neighbor that corresponds already to a region, to the red region.

And it has another neighbor that corresponds to another region, which means that in reality these two regions are connected to each other via our point A. So what will happen here? Well, we would actually merge these two regions and assign A to the merged region and delete one of these regions.

So that would be the outcome. Two regions have been merged. Well, in this case, the pixels would get assigned to the existing region. Here we would have a pixel that has two papers that are not yet assigned to any region, so we initialize a new region.

And then we move on as earlier and until we are finished, and the result is the label in which we want to have. So we have these five different cases that can occur here, and as we may have regions that look like this.

So in this case, we would initialize a region here. We go through the picture overall, so we would initialize another region coming here.

We would realize that these two regions actually belong together, so we can merge them. And then we move on as this graph earlier. This is what it looks like here. It looks like this piece over here. So this is a very simple way of generating these label images.