Dear net,
One way can be to examine the average “amount of noise” inside each tile.
You could have several line segments within each tile (from edge to edge, diagonals…) and use improfile to get the pixel value. Then, you could get the standard deviation (or variance) to see how spread out your values are.
You can see in the images below profile from edge to edge for a flat and rough tile, respectively. Pixel values are from the red channel. For smooth tiles, provided homogeneous illumination and proper recording of the picture, the values will be pretty close, resulting in a small deviation. On the other hand, for the rough tile due to contrast between shadow and elevated features, the standard deviation value is higher.
I would recommend to exclude the first few pixel value near the border from such calculations, because the illumination angle give to smooth tiles a very bright (high values) corner. [PS: This could actually could somehow also be exploited to see the difference from the rough tiles]
Best
