'''
Methods for filtering
We cover implementations for common use cases. One can fall back to the
low level OpenCV API for more flexibility if needed.
'''
import numpy as np
from skimage import img_as_float
import cv2
from cr import vision
from cr.vision.errors import check_ndim, check_nchannels
[docs]def filter_2d(image, kernel):
'''Filters an image retaining its depth
:param image: Input image
:type image: array_like
:param kernel: 2D convolutional Kernel to apply on image
:type kernel: array_like
:return: Filtered image
:rtype: ndarray
'''
# we set ddepth=-1 in the folllowing call
return cv2.filter2D(image, -1, kernel)
[docs]def gaussian_blur(image, kernel_size=5, sigma=0):
'''
Filters an image using Gaussian filter of a given kernel size
:param image: Input image
:type image: array_like
:param kernel_size: Size of Gaussian kernel, defaults to 5
:type kernel_size: int, optional
:param sigma: Kernel standard deviation
:type sigma: double, optional
:return: Gaussian blurred image
:rtype: ndarray
Remarks
* We assume that the filter is symmetric.
'''
kernel_size = (kernel_size, kernel_size)
return cv2.GaussianBlur(image, kernel_size, sigma)
[docs]def laplacian(image, ddepth=cv2.CV_32F):
'''Computes Laplacian of gray scale images '''
check_ndim(image.ndim, 2, 3)
if image.ndim == 3:
_, _, channels = image.shape
check_nchannels(1, channels)
return cv2.Laplacian(image, ddepth)
[docs]def sobel_x(image, ddepth=cv2.CV_32F, ksize=3):
'''Computes sobel derivative in x direction'''
return cv2.Sobel(image, ddepth, 1, 0, ksize=ksize)
[docs]def sobel_y(image, ddepth=cv2.CV_32F, ksize=3):
'''Computes sobel derivative in y direction'''
return cv2.Sobel(image, ddepth, 0, 1, ksize=ksize)
[docs]def sobel_energy_l1(image):
'''Computes the energy matrix for the image using sobel gradients by taking their absolute sums'''
image = vision.bgr_to_gray(image)
# bring image to 0,1 range
image = img_as_float(image)
x_g = sobel_x(image, ddepth=cv2.CV_64F)
y_g = sobel_x(image, ddepth=cv2.CV_64F)
energy = np.absolute(x_g) + np.absolute(y_g)
return energy
[docs]def sobel_energy_l2(image):
'''Computes the energy matrix for the image using sobel gradients by computing the gradient magnitude'''
image = vision.bgr_to_gray(image)
# bring image to 0,1 range
image = img_as_float(image)
x_g = sobel_x(image)
y_g = sobel_x(image)
energy = np.sqrt(x_g*x_g + y_g*y_g)
return energy