【Katie Bouman】How to take a picture of a black hole (Part 4)

20190416_165013.jpg
한국어 日本語

Reconstruction and Sketch Artists

Different types of images have very distinct features. We can easily tell the difference between black hole simulation images and images we take every day here on Earth. We need a way to tell our algorithms what images look like without imposing one type of image's features too much.

One way we can try to get around this is by imposing the features of different kinds of images and seeing how the type of image we assume affects our reconstructions. If all images' types produce a very similar-looking image, then we can start to become more confident that the image assumptions we're making are not biasing this picture that much. 

This is a little bit like giving the same description to three different sketch artists from all around the world. If they all produce a very similar-looking face, then we can start to become confident that they're not imposing their own cultural biases on the drawings.

20190416_164750.jpgㅇ
Differens Puzzle Pieces 'Black Hole' 'Astronomical' 'Everyday'

One way we can try to impose different image features is by using pieces of existing images. So we take a large collection of images, and we break them down into their little image patches. We then can treat each image patch a little bit like pieces of a puzzle. And we use commonly seen puzzle pieces to piece together an image that also fits our telescope measurements.

Different types of images have very distinctive sets of puzzle pieces. So what happens when we take the same data but we use different sets of puzzle pieces to reconstruct the image? Let's first start with black hole image simulation puzzle pieces. OK, this looks reasonable. This looks like what we expect a black hole to look like.

20190416_164711.jpg
ㅇ
But did we just get it because we just fed it little pieces of black hole simulation images? Let's try another set of puzzle pieces from astronomical, non-black hole objects. OK, we get a similar-looking image.

And then how about pieces from everyday images, like the images you take with your own personal camera? Great, we see the same image. When we get the same image from all different sets of puzzle pieces, then we can start to become more confident that the image assumptions we're making aren't biasing the final image we get too much.

１　　２　　３　　４　　５