[DCH3416]

From what I can tell it's using a neural network to derive an image from the interference patterns of light.

I imagine you could do this using a standard computational model, it would just be very intensive. So I guess it would be 'enhanced' in the same way a JPEG stores an image in a lossy format.

[d--b]

My question was more about what it was that records the patterns of light.

[DCH3416]

The nature article shows some sort of cmos like sensor with a surface made of pegs which seem to be conveniently close in size to the wavelengths of visible light. That passes through some sort of meta optic which presumably measures the diffraction off the sensor surface. Both sensor and "meta optic" data combined and extrapolated to form an image.

It's quite a clever way of designing a "lens" like that. Because you can generate an image from practically a flat surface. Of course the output image is "calculated" instead of just bending light through a series of glass lenses.