腐蚀算法在图像清晰度很差的情况下，效果不佳



In [12]:

    
def corrode(img,o):   
    size = img.size
    nimg = Image.new('1',size,255)
    mat = img.load()
    nmat = nimg.load()
    for i in range(o[0],size[0]-o[1]):
        for j in range(o[2],size[1]-o[3]):
            nmat[i,j] = 0
            for k in range(i - o[0],i + o[1] + 1):
                for l in range(j - o[2],j + o[3] + 1):
                    if mat[k,l] == 255:
                        nmat[i,j] = 255
    return nimg
def expand(image,o):
    size = img.size
    nimg = Image.new('1',size,255)
    mat = img.load()
    nmat = nimg.load()
    for i in range(o[0],size[0]-o[1]):
        for j in range(o[2],size[1]-o[3]):
            if mat[i,j] == 0:
                for k in range(i - o[0],i + o[1] + 1):
                    for l in range(j - o[2],j + o[3] + 1):
                        nmat[i,j] = 0
    return nimg
def gray(img,k):
    size = img.size
    nimg = Image.new('L',size,255)
    mat = img.load()
    nmat = nimg.load()
    for i in range(size[0]):
        for j in range(size[1]):
            pix = mat[i,j]
            nmat[i,j] = pix[0] * k[0] + pix[1] * k[1] + pix[2] * k[2]
#            if pix[0] * k[0] + pix[1] * k[1] + pix[2] * k[2] >= 128:
#                nmat[i,j] = 255
#            else:
#                nmat[i,j] = 0
    return nimg



In [2]:

    
from PIL import Image
im = imread('CheckCode.gif')
img = Image.fromarray(im)
#img = img.convert('1')
#nimg = gray(img,(0.3,0.59,0.11))
nimg = gray(img,(1,1,-1))
img = nimg.convert('1')
o = (0,1,0,1)
n1 = corrode(img,o)
n2 = expand(img,o)
n3 = expand(n1,o)
n4 = corrode(n2,o)
imshow(nimg)









    Out[2]:





<matplotlib.image.AxesImage at 0x7f51fc1ae790>



In [28]:

    
imshow(n1)









    Out[28]:





<matplotlib.image.AxesImage at 0x7f92417cd2d0>



In [29]:

    
imshow(n2)









    Out[29]:





<matplotlib.image.AxesImage at 0x7f9241733ad0>



In [30]:

    
imshow(n3)









    Out[30]:





<matplotlib.image.AxesImage at 0x7f924171b650>



In [31]:

    
imshow(n4)









    Out[31]:





<matplotlib.image.AxesImage at 0x7f92416871d0>

使用K-Means算法



In [19]:

    
from sklearn import cluster, datasets
from PIL import Image
im = imread('CheckCode.gif')
size = (len(im[0]),len(im))
l = []
for i in range(size[1]):
    for j in range(size[0]):
        l.append(list(im[i][j]) + [i] + [j])
iris = datasets.load_iris()
k_means = cluster.KMeans(n_clusters = 2)
k_means.fit(l)
img = Image.new('1',size)
mat = img.load()
for i in range(size[1]):
    for j in range(size[0]):
        mat[j,i] = k_means.labels_[i*size[0]+j] * 255
imshow(img)









    Out[19]:





<matplotlib.image.AxesImage at 0x7f5a9c04fb50>

结果，目前已知的最好效果是这样实现的



In [25]:

    
from PIL import Image
import urllib.request
from io import BytesIO

f = urllib.request.urlopen('http://xk.suda.edu.cn/CheckCode.aspx')
s = BytesIO(f.read())
im = imread(s,format='gif')
subplot(2,1,1)
imshow(Image.fromarray(im))
size = (len(im[0]),len(im))
for i in range(size[1]):
    for j in range(size[0]):
        if tuple(im[i,j]) == (0,0,153,255):
            im[i][j] = [0,0,0,255]
        else:
             im[i][j] = [255,255,255,255]
            
img = Image.fromarray(im)
subplot(2,1,2)
imshow(img)









    Out[25]:





<matplotlib.image.AxesImage at 0x7f7bb07cfbd0>

哭

下面尝试面积投影法来分割



In [26]:

    
def shadow(mat,shape):
    l = []
    for i in range(shape[1]):
        c = 0
        for j in range(shape[0]):
            if mat[j,i][0] == 0:
                c = c + 1
        l.append(c)
    return l
l = shadow(im,im.shape)
sub = []
last = 0
for i in range(len(l)):
    if l[i] == 0:
        if i - last > 1:
            sub.append(Image.fromarray(im[:,last:i+1]).convert('1'))
        last = i
c = 1
for i in sub:
    subplot(1,4,c)
    c = c + 1
    imshow(i)

等宽分割效果也不好

下面使用卡钳算法摆正图片



In [27]:

    
def shadow(mat,shape):
    l = []
    for i in range(shape[0]):
        c = 0
        for j in range(shape[1]):
            if  mat[i,j] == 0:
                c = c + 1
        l.append(c)
    return l
font = []
c = 1
for img in sub:
    subplot(2,4,c)
    imshow(img)
    minl = 100
    img = img.point(lambda i:255-i)
    for i in range(-10,10):
        t = img.rotate(i,expand=1)
        t = t.point(lambda i:255-i)
        size = t.size
        l = shadow(t.load(),size)
        for i in range(size[0]):
            if l[i] > 0:
                break
        for j in range(size[0] - 1,-1,-1):
            if l[j] > 0:
                break
        if j - i < minl:
            minp = t
            minl = j - i
    font.append(minp)
    subplot(2,4,c+4)
    c = c + 1
    imshow(minp)

最后OCR



In [62]:

    
import pyocr
tool = pyocr.get_available_tools()[0]
for i in font:
    print(tool.image_to_string(i))

直接使用pyocr会有问题，这里使用命令行执行



In [28]:

    
import os
res = []
for i in font:
    i.convert('RGB').save('tmp.gif')
    os.system('tesseract tmp.gif tmp -psm 10 tesseract_config')
    with open('tmp.txt','r') as f:
        res.append(f.read()[0])
for i in res:
    print(i)