Does Pyvips have uniformly sampled noise?

[stephandooper]

Hi,

I'm trying to create a function in Pyvips that perform elastic deformation transformations. I managed to get the function itself to work, however the algorithm I am trying to implement expects uniform sampled noise in the [0,1] range[1]. My question is if there is such a function in Pyvips, since I managed to only find gaussian, perlin, and worley noise.

The way I circumvented this is by using an inverse box-muller transformation on two independent Gaussians, but I recon that this is not the most efficient way to do this.

P.s.: is there a way to set the seed globally for the noise functions, or is defined per function?

def elastic_deformation(image, sigma=16, alpha=1):
    width, height, channels = image.width, image.height, image.bands

    # Create a random displacement field, pyvips does not have uniform
    # instead, use a Gaussian and convert using Box-Muller inverse
    z1 = pyvips.Image.gaussnoise(width, height, sigma=1.0, mean=0.0, seed=1)
    z2 = pyvips.Image.gaussnoise(width, height, sigma=1.0, mean=0.0, seed=1)

    # Compute box-muller inverse to get approximate uniform values
    dx = (-(z1*z1 + z2*z2) / 2).exp()
    dx = (2 * dx - 1).gaussblur(sigma) * alpha

    dy = (z1 / z2).atan()
    dy = (2 * dy - 1).gaussblur(sigma) * alpha


    grid = pyvips.Image.xyz(image.width, image.height)
    map = grid + dx.bandjoin([dy])

    image = image.mapim(map, interpolate=pyvips.Interpolate.new('bilinear'), background=[255, 255, 255])
    return image

[1] Simard2003] Simard, Steinkraus and Platt, "Best Practices for Convolutional Neural Networks applied to Visual Document Analysis", in Proc. of the International Conference on Document Analysis and Recognition, 2003

[jcupitt]

Hi @stephandooper,

You're right, there's no simple uniform noise generator, I suppose I've never needed one. Let's tag this as an enhancement for 8.16.

I can see a global seed might be useful too. You can call g_random_set_seed():

https://developer-old.gnome.org/glib/stable/glib-Random-Numbers.html#g-random-set-seed

But of course that's not exposed by pyvips.

[jcupitt]

I added an enhancement issue, let's continue any discussion there.