如何用3x3内核平滑图像

Question

我试图平滑一个图像，通过循环遍历它的像素，计算一个3x3补丁的平均值，然后将平均值应用到这个补丁中的所有9个像素。

代码：

    import matplotlib.pyplot as plt
    import numpy as np
    import cv2 as cv
    from PIL import Image


    # 1. Load Image
    name = 'ebay.png'
    img = cv.imread(name) #import image

    h, w = img.shape[:2]


    # 2. Smooth with kernel size 3

    for y in range(0, w, 3):
        for x in range(0, h, 3):     
            px1 = img[x][y] #0/0
            px2 = img[x][y+1] #0/1
            px3 = img[x][y+2] #0/2
            px4 = img[x+1][y] #1/0
            px5 = img[x+1][y+1] #1/1
            px6 = img[x+1][y+2] #1/2
            px7 = img[x+2][y] #2/0
            px8 = img[x+2][y+1] #2/1
            px9 = img[x+2][y+2] #2/2

            average = np.average(px1 + px2 + px3 + px4 + px5 + px6 + px7 + px8 + px9)    

            img[x][y] =  average  #0/0
            img[x][y+1] = average   #0/1
            img[x][y+2] = average   #0/2
            img[x+1][y] = average   #1/0
            img[x+1][y+1] = average   #1/1
            img[x+1][y+2] = average  #1/2
            img[x+2][y] = average  #2/0
            img[x+2][y+1] = average   #2/1
            img[x+2][y+2] = average  #2/2

# 3. Transform the resulting image into pgm format and save result        

new_image = Image.fromarray(img)
new_image.save('new.png')

# 4. Show image

new_image.show()

However this just makes my new image just very pixely and not smooth at all. 

我假设我在这里做错了什么：

average = np.average(px1 + px2 + px3 + px4 + px5 + px6 + px7 + px8 + px9)

因为，当我只使用px5作为平均值时，新的图像看起来要好得多(但仍然不是很流畅)。请参阅下图：

原始图像：

​

我的代码现在所做的是：

​

结果当我使用px5作为平均值时：

​

结果当添加所有px并除以9：

​

Answer 1

因此，我在这里有两个问题，我能够弄清楚，这要感谢@Ernie Yang和@Cris Luengo。非常感谢您的帮助！

1)我的平均计算的问题是，它是溢出的，我是在总结像素值。这就是为什么结果看起来很奇怪，因为它被包围了。所以我不得不改变：

average = np.average(px1 + px2 + px3 + px4 + px5 + px6 + px7 + px8 + px9)

至：

average = px1/9. + px2/9. + px3/9. + px4/9. + px5/9. + px6/9. + px7/9. + px8/9. + px9/9.

2)然而，这并没有使我的图像平滑，因为我只是将平均值分配给补丁中的所有9个像素。因此，这导致图片被像素化，而不是平滑。因此，我不得不将平均值的结果只写到中间像素，而不是写到邻里所有的3x3像素。我还必须把它写到一个单独的输出映像中。您无法在适当位置执行此操作，因为它将影响后续像素的结果。

正确的代码示例：

    import matplotlib.pyplot as plt
    import numpy as np
    import cv2 as cv
    from PIL import Image
    import scipy.ndimage as ndimage
    from scipy.ndimage.filters import gaussian_filter


    # 1. Load Image
    name = 'ebay.png'
    img = cv.imread(name) #import image


    h, w = img.shape[:2]
    smoothedImage = cv.imread(name) #initialize second image


    # 2. Smooth with with kernel size 3

    for y in range(0, w-2):
        for x in range(0, h-2):     
            px1 = img[x][y] #0/0
            px2 = img[x][y+1] #0/1
            px3 = img[x][y+2] #0/2
            px4 = img[x+1][y] #1/0
            px5 = img[x+1][y+1] #1/1
            px6 = img[x+1][y+2] #1/2
            px7 = img[x+2][y] #2/0
            px8 = img[x+2][y+1] #2/1
            px9 = img[x+2][y+2] #2/2

            average = px1/9. + px2/9. + px3/9. + px4/9. + px5/9. + px6/9. + px7/9. + px8/9. + px9/9.

            smoothedImage[x+1][y+1] = average   #1/1

    # 3. Transform the resulting image into pgm format and save result        

    new_image = Image.fromarray(smoothedImage)
    new_image.save('new.png')

    # 4. Show image

    new_image.show()

原始图像：

​

平滑图像：

​

编辑：

嘿，伙计们，我下午午睡回来了。我有很多有趣的想法，下面是我改进的代码：

import matplotlib.pyplot as plt
import numpy as np
import cv2 as cv
from PIL import Image
import scipy.ndimage as ndimage
from scipy.ndimage.filters import gaussian_filter


# 1. Load Image
name = 'ebay.png'
img = cv.imread(name) #import image
h, w = img.shape[:2]
kernel = 5
radius = (kernel-1)/2

img2 = np.zeros((h, w, 3), dtype = 'uint8') #new image to paint on

    def pxIsInImgRange(x, y):
        if (0<=x) and (x < w): 
                if (0<=y) and (y < h):
                    return True
        return False

    # 2. Smoothing the shit out

    for x in range (-radius, w+radius):
        for y in range (-radius, h+radius):

            if pxIsInImgRange(x,y): 

                    px = 0

                    for vx2 in range (-radius, radius+1):
                        for vy2 in range (-radius, radius+1):
                            x2 = x + vx2
                            y2 = y + vy2
                            if pxIsInImgRange(x2,y2):

                                px = px + (img[y2][x2]/float((kernel*kernel)))
                            else:
                                px = px + 0 


                    img2[y][x] = px

    # 3. Save image                

    new_image = Image.fromarray(img2)
    new_image.save('new.png')

    # 4. Show image

    new_image.show()

内核为5的新结果：

​

​

Answer 2

average = np.average(px1 + px2 + px3 + px4 + px5 + px6 + px7 + px8 + px9)

是对数值的实际求和，而不是平均值，

average = (px1 + px2 + px3 + px4 + px5 + px6 + px7 + px8 + px9)/9

应该给你你想要的。

对于这样的任务，scipy.signal.convolve2d也是最好的工具。参见下面的文档和示例。

https://www.google.com/search?q=scipy.signal.convolve2d&oq=scipy.signal.convolve2d&aqs=chrome..69i57j69i60&sourceid=chrome&ie=UTF-8