About scipy Gaussian filter
scipy.ndimage.gaussian_filter [reference]
usage.py
from scipy.ndimage import gaussian_filter
blur_array = gaussian_filter(input, sigma)
As of v1.5.4, the arguments look like this.
| argument | Explanation | Default |
|---|---|---|
| input | It is OK if it is in array format. | Mandatory |
| sigma | Standard deviation of Gaussian distribution, degree of blurring. Can be set for each axis. (Example)In 3D sigma=[3,6,9] |
Mandatory |
| order | Derivative rank, if it is 1 or more, order the filter. Can be set for each axis. | 0 |
| output | Output data format. | Same as input |
| mode | Behavior when applying a filter at the edge, described later. (type) ‘reflect’ ``‘constant’``‘nearest’ ``‘mirror’``‘wrap’ |
‘reflect’ |
| cval | A value outside the input array. mode=Applicable when ‘constant’. | 0.0 |
| truncate | The scope of application of the filter. Filter radius=int(sigma*truncate+0.5) | 4.0 |
- When a color image (HEGHT, WIDTH, 3) is input, it is also smoothed on the third axis (color channel direction), so it is necessary to set
sigma = [n, n, 0]. For image formats, cv2 and skimage are easy. On the contrary, scipy has the advantage that it can be applied to any dimensional tensor (?)
About sigma
・ Scalar values are the same for all axes $ \ sigma $. ・ If it is a sequence (number of elements = input dimension), $ \ sigma $ is specified for each axis.
Result example
| Color image Color smoothing |
Color image | Grayscale image | Smoothing only in the X direction | Smoothing only in the Y direction |
|---|---|---|---|---|
 |
 |
 |
 |
 |
About mode
-How to handle values outside the edge (edge of the array).
・ The default is ‘reflect’
-The behavior when the input sequence is abcd is described in the table.
| mode= | Explanation | Out of array left | Input array | Out of array right |
|---|---|---|---|---|
| ‘reflect’ | Reflected around the edge.(Edges are also reflected) | d c b a | a b c d | d c b a |
| ‘constant’ | constantcval， Default is 0 |
k k k k | a b c d | k k k k |
| ‘nearest’ | The value of the closest edge. | a a a a | a b c d | d d d d |
| ‘mirror’ | Reflected around the edge.(Edges do not reflect) | d c b | a b c d | c b a |
| ‘wrap’ | Wrap around the opposite edge | a b c d | a b c d | a b c d |
About truncate
・ Smoothing range ・ Default is 4 -If you check the code, probably the filter radius = int (sigma * truncate + 0.5) [code]
Result example
If truncate is 4 or more, there is almost no difference, so there seems to be no need to mess with it.
