This made me think that maybe it's doable via machine learning. Upscale the image to high resolution with super resolution technology, like waifu2x, then pixelate it back to 48x48 with something like https://github.com/sedthh/pyxelate.
I've never tried such process, don't know if it works. But it's certainly better than simply up-scaling the image. But still, redrawing is unavoidable for the best quality.