Slightly boringly, this very similar to my last post – but it’s also something useful that you may want to know, and that I’ll probably forget if I don’t write it down somewhere.
Basically, scipy.ndimage.filters.convolve is about twice as fast as scipy.signal.convolve2d.
I run convolutions a lot on satellite images, and Landsat images are around 8000 x 8000 pixels. Using a random 8000 x 8000 pixel image, with a 3 x 3 kernel (a size I often use), I find that convolve2d takes 2.9 seconds, whereas convolve takes 1.1 seconds – a useful speedup, particularly if you’re running a loop of various convolutions. The same sort of speed-up can be seen with larger kernel sizes, for example, a 9 x 9 kernel takes 14.7 seconds and 6.8 seconds – an even more useful speed-up.
When I switched my code from convolve2d to convolve, I had to do a bit of playing around to get the various options set to ensure that I got the same results. For reference, the following two lines of code are equivalent:
res1 = convolve2d(img, kernel, mode='same') res2 = convolve(img, kernel, mode='constant', cval=0.0)
Interestingly these functions are both within scipy, so there must be a reason for them both to exist. Can anyone enlighten me? My naive view is that they could be merged – or if not, then a note could be added to the documentation saying that convolve is far faster!