使用Python和Numpy将RGB图像与自定义邻居内核卷积在一起

Question

我正在尝试实现一种算法来验证RGB图像的4个邻居（上，下，左和右）像素，如果所有像素RGB值都相等，我将输出图像中的像素标记为1，否则它将为0.非矢量化实现是：

def set_border_interior(img):
  img_rows = img.shape[0]
  img_cols = img.shape[1]
  res = np.zeros((img_rows,img_cols))
  for row in xrange(1,img_rows-1):
      for col in xrange(1,img_cols-1):
          data_b = set()
          data_g = set()
          data_r = set()
          up = row - 1
          down = row + 1
          left = col - 1
          right = col + 1

          data_b.add(img.item(row,col,0))
          data_g.add(img.item(row,col,1))
          data_r.add(img.item(row,col,2))

          data_b.add(img.item(up,col,0))
          data_g.add(img.item(up,col,1))
          data_r.add(img.item(up,col,2))

          data_b.add(img.item(down,col,0))
          data_g.add(img.item(down,col,1))
          data_r.add(img.item(down,col,2))

          data_b.add(img.item(row,left,0))
          data_g.add(img.item(row,left,1))
          data_r.add(img.item(row,left,2))

          data_b.add(img.item(row,right,0))
          data_g.add(img.item(row,right,1))
          data_r.add(img.item(row,right,2))

          if (len(data_b) == 1) and (len(data_g) == 1) and (len(data_r) == 1):
              res.itemset(row,col, False)
          else:
              res.itemset(row,col, True)
  return res

这种非矢量化方式，但它确实很慢（甚至使用img.item读取数据和img.itemset来设置新值）。有没有更好的方法在Numpy（或scipy）中实现这个？

Answer 1

将边框放在一边，无论如何你的功能都没有明确定义，你可以做到以下几点：

import numpy as np
import matplotlib.pyplot as plt

rows, cols = 480, 640
rgb_img = np.zeros((rows, cols, 3), dtype=np.uint8)

rgb_img[:rows//2, :cols//2] = 255

center_slice = rgb_img[1:-1, 1:-1]
left_slice = rgb_img[1:-1, :-2]
right_slice = rgb_img[1:-1, 2:]
up_slice = rgb_img[:-2, 1:-1]
down_slice = rgb_img[2:, 1:-1]

all_equal = (np.all(center_slice == left_slice, axis=-1) &
             np.all(center_slice == right_slice, axis=-1) &
             np.all(center_slice == up_slice, axis=-1) &
             np.all(center_slice == down_slice, axis=-1))

plt.subplot(211)
plt.imshow(rgb_img, interpolation='nearest')
plt.subplot(212)
plt.imshow(all_equal, interpolation='nearest')
plt.show()