That's a more recent trend. Until the late 2000s, computer vision was basically combining machine learning techniques with image processing: Design filters to extract features, and slap them into a classifier. Naturally, lots of Fourier, wavelets, and other weird bases. Very different times.