基于连通域字符分割的流程_基于连通域的字符分割

要做文字图像识别，首先得分割字符。之前写过连通域的分割算法，这次就用该算法来分割字符。假设有如下图片，

首先基于连通域进行分割：

import numpy as npimport cv2from matplotlib import pyplot as pltclass ImgSplit():    def __init__(self,path):        self.img0 = cv2.imread(path,cv2.IMREAD_GRAYSCALE)#读成灰度图        img = self.preProcess(self.img0,120)        self.domains = self.connectedDomain(img)    def preProcess(self,img0, thresh= 50):        img = img0.copy()        #二值化，同时黑白颠倒        img[np.where(img < thresh)] = 1        img[np.where(img >= thresh)] = 0        return img             def connectedDomain(self, img):        h,w = img.shape        # 4-connected domain        domain =dict()        label = -1 # 用于标记连通域ID        #i==0 # 第一行        flag = True #上一个像素是否为0？        for j in range(w):            if img[0,j]==1:                if flag: #上一个元素是0则新建连通域                    label += 1                    domain[label] = set()                domain[label].add((0,j))                flag = False            else:                flag = True        for i in range(1,h):#第1行外的其它行            flag = True            for j in range(w):                if img[i,j] == 1:                    for key in domain:                        if (i-1,j) in domain[key]:#上一行对应元素是否属于某个连通域？                            domain[key].add((i,j))#记录像素坐标                                                        if not flag:#左边的元素为1                                if last_label != key:                                    domain[key]=domain[key].union(domain[last_label]) #合并                                    del domain[last_label]                                                                last_label = key                                break                    else:#不属于上行已有的连通域                        if flag:                            label += 1                            domain[label] = set()                            domain[label].add((i,j))                            last_label = label                        else:                            domain[last_label].add((i,j))                                        flag = False                else:                    flag = True        return domain.values()#imgSplit = ImgSplit("test.PNG")imgSplit = ImgSplit("3.PNG")domains = imgSplit.domains#按面积从大到小排序#descended = sorted(domains,key = len,reverse=True)#从上到下排列(按非零像素的平均x坐标)#up2down  = sorted(domains,key = lambda g: sum([coord[1] for coord in g])/ float(len(g)))#从左到右排序 (按非零像素的平均y坐标)left2right  = sorted(domains,key = lambda g: sum([coord[1] for coord in g])/ float(len(g)))’‘’for i,domain in enumerate(left2right):    z = np.zeros((imgSplit.img0.shape))    for x,y in domain:        z[x,y] =1    plt.subplot(10,12,i+1)    plt.imshow(z)plt.show()‘’‘

结果分割后的碎片太多，有一百多块。

我们把左右靠的近的碎片合并(这里假设了字符左右排列)：

yMeans = []yMins = []yMaxs =[]for domain in left2right:    Y = [item[1] for item in domain]    yMins.append(min(Y))    yMaxs.append(max(Y))    yMeans.append(sum(Y)/float(len(Y)))    #合并上下型不该分开的连通域for i in range(len(yMeans)-1,0,-1):    buffer = 0.21 #buffer rate    if yMins[i-1]-buffer*(yMaxs[i-1]-yMins[i-1]) < yMeans[i]< yMaxs[i-1]+ buffer*(yMaxs[i-1]-yMins[i-1]) \       or yMins[i] -buffer*(yMaxs[i]-yMins[i]) < yMeans[i-1]< yMaxs[i]+buffer*(yMaxs[i]-yMins[i]):        left2right[i-1] =left2right[i-1].union(left2right[i])        yMaxs[i-1]= max(yMaxs[i], yMaxs[i-1])        yMins[i-1] = min(yMins[i], yMins[i-1])        del left2right[i]        del yMeans[i]        del yMins[i]        del yMaxs[i]        for i,domain in enumerate(left2right):            z = np.zeros((imgSplit.img0.shape))    for x,y in domain:        z[x,y] =1    plt.subplot(5,5,i+1)    plt.imshow(z)plt.show()

我们得到了16块带黑色背景的字符块儿：

最后把字符中心区域抠出来即可：

xMins = []xMaxs = []for domain in left2right:     X = [item[0] for item in domain]    xMins.append(min(X))    xMaxs.append(max(X))heightMax = max(xMaxs) - min(xMins)+1widthMax = int(42/26.0*heightMax) #42/26：训练集图像的长宽比for i,domain in enumerate(left2right):    z = np.zeros((imgSplit.img0.shape))    for x,y in domain:        z[x,y] =1    z = z[min(xMins):max(xMaxs)+1, yMins[i]:yMaxs[i]+1]    plt.subplot(4,4,i+1)    plt.imshow(z)plt.show()

再来一段英文的：

分割得(空格被忽略了)：

这里只是提供一种思路，因为这种先拆得稀碎再合并的算法效率不高，应该适当地修改分割算法。