You're right with its common application. But the Hough transformation isn't binary at all. The principle of the Hough transform is that you extract prominent points from a structure or object and sum up the occurrences in another space, the Hough space. The idea behind is that parametric shapes in an image are detected when looking for accumulation points in the parametric space. See Kittler et al. "A survey of the hough transform" computer vision, graphics, and image processing vol 44, No1, pp87-116, and Edwards et al. "Generalized Hough transform for natural shapes", pattern recognition letters Vol 18, No 5

So you can use thresholding to get binary images as input to the Hough transform, or you can calculate 5x5 up to nxn gradient xy-vectors from gray scale images, calculate from that magnitude/angle, or use difference of Gaussians(DOG) instead of gradient and finally apply thinning to the magnitude as input to Hough transform.

Since building gradients form noisy images is amplifying the noise, you can superpose several instances of your EBSD pattern files when they are static. The signal/noise ratio increases proportional to sqrt(N), when N is the number of files. You might get brilliant looking images as input files.

 Maybe, I did nor expressed myself accurately. Hopefully the attached image explains it better. There a common pattern (160x120 pixels) from a measurement is shown (Fayalite, Fe2SiO4, orthorhombic) which is Hough-transformed (120x120 pixels). The same pattern is also binarized and then Hough-transformed (what is -as far as I know - the common procedure in all commercial EBSD software systems). The result looks a bit different, but after a bit rescaling both Hough transforms don't look that much different anymore. The question is: Has the Hough transform - which is then perhaps called Radon transform - applied on grey-scaled images any real advantage, and when, how big is it? Does it make sense to spend this time for a "more accurate" orientation determination?