The "numpy" Sobel code is not that good, unfortunately - all the iteration is done in Python, so there is not much benefit from involving numpy. If one would use say scipy.convolve2d on a numpy.array, it would be much faster.