You could build a solution off of this:

cv2.matchTemplate at one scale → works only if sizes match; breaks with scale/rotation.

Come up with a range of possible scales and a discretization of rotations. Compute the worst case for each direction of resizing and for each scale, scale the template to that scale. Then, for each rotation, rotate the template. Now, run cv2.matchTemplate for each of the possible templates.

Or go crazy and do some sort of equivariant steerable CNN pyramid network, and train it over all your symbols and example images.