[_0ffh]

Can surely be done with the original neural style transfer algorithm. You'd just have to use a soft mask over the picture when calculating the content loss function.

Don't know if there's an analog way for fast style transfer.