[VP] Sobel, Scharr, Prewitt, Roberts四种滤波器的对比

---

所用图像如下，使用时请自行更换Image：

代码

import numpy as np
from scipy.ndimage import rotate
import os
from PIL import Image
import math
import cv2
import matplotlib.pyplot as plt
%matplotlib inline

#Sobel Kernel
sobelH = np.array([[1, 0, -1], 
                   [2, 0, -2], 
                   [1, 0, -1]])
sobelV = sobelH.T

#Scharr Kernel
scharrH = np.array([[3, 0, -3], 
                   [10, 0, -10], 
                   [3, 0, -3]])
scharrV = scharrH.T

#Prewitt Kernel
prewittH = np.array([[1, 0, -1], 
                    [1, 0, -1], 
                    [1, 0, -1]])
prewittV = prewittH.T

#roberts Kernel
robertsH = np.array([[0, 1],  
                    [-1, 0]])
robertsV = np.array([[1, 0],  
                    [0, -1]])


im1SH = cv2.filter2D(im1a, kernel = sobelH, ddepth = -1)
im1SV = cv2.filter2D(im1a, kernel = sobelV, ddepth = -1)
im2SH = cv2.filter2D(im2a, kernel = sobelH, ddepth = -1)
im2SV = cv2.filter2D(im2a, kernel = sobelV, ddepth = -1)

im1ScH = cv2.filter2D(im1a, kernel = scharrH, ddepth = -1)
im1ScV = cv2.filter2D(im1a, kernel = scharrV, ddepth = -1)
im2ScH = cv2.filter2D(im2a, kernel = scharrH, ddepth = -1)
im2ScV = cv2.filter2D(im2a, kernel = scharrV, ddepth = -1)

im1PH = cv2.filter2D(im1a, kernel = prewittH, ddepth = -1)
im1PV = cv2.filter2D(im1a, kernel = prewittV, ddepth = -1)
im2PH = cv2.filter2D(im2a, kernel = prewittH, ddepth = -1)
im2PV = cv2.filter2D(im2a, kernel = prewittV, ddepth = -1)

im1RH = cv2.filter2D(im1a, kernel = robertsH, ddepth = -1)
im1RV = cv2.filter2D(im1a, kernel = robertsV, ddepth = -1)
im2RH = cv2.filter2D(im2a, kernel = robertsH, ddepth = -1)
im2RV = cv2.filter2D(im2a, kernel = robertsV, ddepth = -1)

fig, ax = plt.subplots(nrows=2, ncols=4, figsize=(20, 10))
plt.axis('off')

ax[0, 0].imshow(im1SH, cmap='gray')
ax[0, 0].set_title("Im1 SobelSH Kernel")
ax[0, 0].axis('off')

ax[0, 1].imshow(im1SV, cmap='gray')
ax[0, 1].set_title("Im1 SobelV Kernel")
ax[0, 1].axis('off')

ax[0, 2].imshow(im1ScH, cmap='gray')
ax[0, 2].set_title("Im1 ScharrH Kernel")
ax[0, 2].axis('off')

ax[0, 3].imshow(im1ScV, cmap='gray')
ax[0, 3].set_title("Im1 ScharrV Kernel")
ax[0, 3].axis('off')

ax[1, 0].imshow(im1PH, cmap='gray')
ax[1, 0].set_title("Im1 PrewittH Kernel")
ax[1, 0].axis('off')

ax[1, 1].imshow(im1PV, cmap='gray')
ax[1, 1].set_title("Im1 PrewittV Kernel")
ax[1, 1].axis('off')

ax[1, 2].imshow(im1RH, cmap='gray')
ax[1, 2].set_title("Im1 RobertsH Kernel")
ax[1, 2].axis('off')

ax[1, 3].imshow(im1RV, cmap='gray')
ax[1, 3].set_title("Im1 RobertsV Kernel")
ax[1, 3].axis('off')


_, ax = plt.subplots(nrows=2, ncols=4, figsize=(18, 5))

ax[0, 0].imshow(im2SH, cmap='gray')
ax[0, 0].set_title("Im2 SobelSH Kernel")
ax[0, 0].axis('off')

ax[0, 1].imshow(im2SV, cmap='gray')
ax[0, 1].set_title("Im2 SobelV Kernel")
ax[0, 1].axis('off')

ax[0, 2].imshow(im2ScH, cmap='gray')
ax[0, 2].set_title("Im2 ScharrH Kernel")
ax[0, 2].axis('off')

ax[0, 3].imshow(im2ScV, cmap='gray')
ax[0, 3].set_title("Im2 ScharrV Kernel")
ax[0, 3].axis('off')

ax[1, 0].imshow(im2PH, cmap='gray')
ax[1, 0].set_title("Im2 PrewittH Kernel")
ax[1, 0].axis('off')

ax[1, 1].imshow(im2PV, cmap='gray')
ax[1, 1].set_title("Im2 PrewittV Kernel")
ax[1, 1].axis('off')

ax[1, 2].imshow(im2RH, cmap='gray')
ax[1, 2].set_title("Im2 RobertsH Kernel")
ax[1, 2].axis('off')

ax[1, 3].imshow(im2RV, cmap='gray')
ax[1, 3].set_title("Im2 RobertsV Kernel")
ax[1, 3].axis('off')

结果分析

• Sobel：索贝尔滤波器对纹理复杂的图形性能较弱，边缘检测精度不够高。在图1的中心，sobel滤波器几乎无法检测到复杂纹理，因此可以看到图1梯度为比较暗的区域。

• Scharr：对于Scharr滤波器，相邻像素的权值较大，因此精度更高，可以计算出比其他滤波器更小的梯度变化。它具有最好的检测精度，这就是为什么使用Scharr滤波器的图像是最亮的。另外我们可以看到主纹理周围有许多额外的渐变区域。

• Prewitt：Prewitt滤波器与索贝尔滤波器相似，但它没有权重的概念，所以结果比索贝尔滤波器更差。

• Roberts：罗伯特滤波器有最好的锐化效果，因为罗伯特滤波器很简单，计算量较少，对细节较敏感。所以边缘更细，整个带有罗伯特滤波器的图像看起来是最暗的。