从文档图像中删除非彩色组件

Question

我需要准备文档图像以进一步识别。我国有严格的规定，文件必须用黑​​色或灰色字体书写。签名和邮票也必须是蓝色的。因此，在当前步骤中，我需要通过删除非灰度彩色元素来删除图章和签名。这是例如文档的片段： 我写的代码转换成灰度级的灰度化成分，并分别非灰度化成分：

#!/usr/bin/env python 
# -*- coding: utf-8 -*- 
import numpy as np 
from skimage import io 
from skimage.color import rgba2rgb, rgb2grey, colorconv 
import matplotlib.pyplot as plt 
def separate_color_regions(im): 
    if im.shape[-1]==4: 
     im=rgba2rgb(im) 
    #representin each color chanel as float in range [0, 1.] 
    im=colorconv._prepare_colorarray(im) 

    #because we wont decompose black color instead of white 
    im=1.-im 

    #calculate gray component using the formula: grey=cos_a*|color|*(1, 1, 1)/sqrt(3), 
    #according to scalar multiplication formula cos_a*|color|=color*(1, 1, 1)/sqrt(3) 
    #here sqrt(3) added since |(1, 1, 1)|=sqrt(3) 
    gray=(im[..., 0]+im[..., 1]+im[..., 2])/3 
    proj=im.copy() 
    for i in range(3): 
     proj[..., i]-=gray 
    def norm(inp): 
     """ 
     convert each color vector to its l2 norm 
     """ 
     t=inp*inp 
     return np.sqrt(np.sum(t, -1)) 
    def mask_nan(inp): 
     where_are_NaNs = np.isnan(inp) 
     inp[where_are_NaNs] = 0. 
     return inp 
    return 1.-mask_nan(norm(proj)/norm(im)), 1.-mask_nan((3.**0.5)*gray/norm(im)) 
im=io.imread('stamps.png') 
fig, (ax1, ax2, ax3) = plt.subplots(nrows=3) 
ax1.imshow(im) 
ax1.axis('off') 
ax1.set_title('Original image') 
nongray, gray=separate_color_regions(im) 
ax2.imshow(nongray, cmap=plt.cm.gray, aspect='auto', interpolation='none') 
ax2.axis('off') 
ax2.set_title('Non-grey_components filter') 
ax3.imshow(gray, cmap=plt.cm.gray, aspect='auto', interpolation='none') 
ax3.axis('off') 
ax3.set_title('Grey component filter') 
fig.tight_layout() 
plt.show() 

这里是它的结果是性能： 您会发现它只能突出显示非灰度对象，但无法将灰度对象与某些灰度对象分离，并且会产生很多噪点。请告诉我是否有更好的解决方案。

Answer 1

与手动seted门槛解决这个问题：

#!/usr/bin/env python 
# -*- coding: utf-8 -*- 
import numpy as np 
from skimage import io 
from skimage.color import rgba2rgb, rgb2grey, colorconv 
import matplotlib.pyplot as plt 
def separate_color_regions(im): 
    if im.shape[-1]==4: 
     im=rgba2rgb(im) 
    #representin each color chanel as float in range [0, 1.] 
    im=colorconv._prepare_colorarray(im) 

    #because we wont decompose black color instead of white 
    im=1.-im 

    #calculate gray component using the formula: grey=cos_a*|color|*(1, 1, 1)/sqrt(3), 
    #according to scalar multiplication formula cos_a*|color|=color*(1, 1, 1)/sqrt(3) 
    #here sqrt(3) added since |(1, 1, 1)|=sqrt(3) 
    gray=(im[..., 0]+im[..., 1]+im[..., 2])/3 
    proj=im.copy() 
    for i in range(3): 
     proj[..., i]-=gray 
    def norm(inp): 
     """ 
     convert each color vector to its l2 norm 
     """ 
     t=inp*inp 
     return np.sqrt(np.sum(t, -1)) 
    def mask_nan(inp): 
     where_are_NaNs = np.isnan(inp) 
     inp[where_are_NaNs] = 0. 
     return inp 
    return 1.-mask_nan(norm(proj)/norm(im)), 1.-mask_nan((3.**0.5)*gray/norm(im)) 
im=io.imread('stamps.png') 
fig, (ax1, ax2, ax3) = plt.subplots(nrows=3) 
ax1.imshow(im) 
ax1.axis('off') 
ax1.set_title('Original image') 
nongray, gray=separate_color_regions(im) 
ax2.imshow(nongray, cmap=plt.cm.gray, aspect='auto', interpolation='none') 
ax2.axis('off') 
ax2.set_title('Non-grey_components filter') 
mask=np.vectorize(lambda d: 0. if d>0.1 else 1.) 
ax3.imshow(1.-(1.-rgb2grey(im))*mask(1.-nongray), cmap=plt.cm.gray, aspect='auto', interpolation='none') 
ax3.axis('off') 
ax3.set_title('Grey component filter') 
fig.tight_layout() 
plt.show()