Initially, my thoughts were going into the direction: "when I train with so random backgrounds but clearly visible symbols on top, the production model will later find laundry symbols on any kind of image the user sends".

First I had the symbols transparent on the background, but it was already super hard for a human to find the symbols, so I thought in the real dataset, the laundry symbols will be always on a single colored background. That's why I started adding random background colors.