CopyPixelsToTensor(TensorSpan<float> dest) API

[vpenades]

I see ImageSharp being used as the backbone for AI image inference more often than not.

In most cases the process involves loading an image with ImageSharp, copying the pixels to a tensor, and running the model.

The tricky interop thing is the "copy pixels to a tensor", seeng how people is doing it , tells me there's a lot of room for improvement.

Preparing an image for a tensor involves two steps:
1- rescale the image to fit into the tensor, sometimes preserving aspect ratio, sometimes not.
2- convert pixels to float, and then apply a std-mean expansion.

Now, the second issue is that the destination is usually a specific data type like the tensors provided by OnnxRuntime or other inference engines.

Microsoft has recently introduced System.Numerics.Tensors.TensorSpan<T> which is readily available, and can provide a middle ground between ImageSharp and any third party inference engine.

So my proposal is to and an extension API that looks like this:

Image.CopyPixelsToTensor(TensorSpan dest, ...)

where additional parameters would tell whether to perform automatic rescaling/fitting, and the std-mean value expasion (which is basically a MAD operation applied to the pixel values)