spkit.conv2d_nan

spkit.conv2d_nan(x, kernel, boundary='fill', fillvalue=0, denormalise_ker=True)

2D Convolution with NaN values

2D Convolution with NaN values

In conventional Convolution funtions, if any of the value in input x or in kernel is NaN (np.nan), then NaN values are propogated and corrupt other values too.

To avoid this, this funtion does the convolution in same fashion as conventional except, it allows NaN values to exists in input, without propogating their effect. It even replaces them.

While computating, it simply ignores the NaN value, as if does not exist, and adjust the computation accordingly.

If No NaN value exist, result is same as conventional convolution

Parameters:
x: 2D-array with NaN values.
kernel: a 2D kernel

  • to use for convolution

  • IMPORTANT NOTE: kernel passed should NOT be normalised. If normalised kernel is used, the results will be very different than conventional convolution.

    For example:

    kernel_unnorm = np.ones([3,3]) kernel_norm = np.ones([3,3])/9

    kernel_unnorm should be passed, not kernel_norm.

  • To de-normalise a kernel, used denorm_kernel

  • or set denormalise_ker=True

denormalise_ker: bool, default=True

  • If True, first de-normalise kernel

boundarystr {‘fill’,’constant’ ,’wrap’, ‘symm’,’symmetric’,’reflect’,}, optional

A flag indicating how to handle boundaries: fill or constant

pad input arrays with fillvalue. (default)

wrap

circular boundary conditions.

symm or ‘symmetric’

symmetrical boundary conditions.

reflect

reflecting boundary conditions.

fillvalue: scalar, optional

Value to fill pad input arrays with. Default is np.nan,

Returns:
y: 2D-arrray of same size as input x with no NaN propogation

See also

conv1d_nan, fill_nans_1d, fill_nans_2d

Examples

>>> #sp.conv2d_nan
>>> import numpy as np
>>> import matplotlib.pyplot as plt
>>> import seaborn as sns
>>> import scipy
>>> import spkit as sp
>>> seed = 2
>>> I = (sp.create_signal_2d(n=10,seed=seed,sg_winlen=3)*10).round(0)
>>> np.random.seed(seed)
>>> r = np.random.rand(*I.shape)
>>> np.random.seed(None)
>>> I[r<0.05] = np.nan
>>> kernel = np.ones([3,3])/9
>>> I_scipy = scipy.signal.convolve2d(I.copy(), kernel,mode='same', boundary='fill',fillvalue=0)
>>> I_spkit = sp.conv2d_nan(I.copy(), kernel, boundary='fill',fillvalue=0)
>>> plt.figure(figsize=(12,4))
>>> plt.subplot(131)
>>> sns.heatmap(I, annot=True,cbar=False,xticklabels='', yticklabels='')
>>> plt.title('Image with NaNs')
>>> plt.subplot(132)
>>> sns.heatmap(I_scipy, annot=True,cbar=False,xticklabels='', yticklabels='')
>>> plt.title('Convolution uisng Scipy \n (scipy.signal.convolve2d)')
>>> plt.subplot(133)
>>> sns.heatmap(I_spkit, annot=True,cbar=False,xticklabels='', yticklabels='')
>>> plt.title('Convolution uisng Spkit \n (sp.conv2d_nan)')
>>> plt.tight_layout()
>>> plt.show()
../../_images/spkit-conv2d_nan-1.png

Examples using spkit.conv2d_nan

Release Highlights for spkit 0.0.9.7

Release Highlights for spkit 0.0.9.7