Python 用的是2.7版本
字符识别用的是libsvm，针对特定的数据进行训练，之后在进行识别，识别率还是比较高的，之前我也用过mnist数据集，结果不是太令人满意
#coding=utf-8
#!/usr/bin/env python
import Image
import cv2
from cv2 import cv
import numpy as np
from pylab import *
import glob
import os
from svmutil import *
'''
分割出图片中的数字，并改变大小，返回值为图片数组 ，如path = 'D:/pic/93.png'
返回的数据格式[[],[],[],.....]
'''
def picSplitResize(path):
image = cv2.imread(path)
''' 获取一张图片四个角的灰度值'''
w,h = cv.GetSize(cv.fromarray(image))
gray = cv2.cvtColor(image,cv2.COLOR_BGR2GRAY)
angleGray = [0] * 4
angleGray[0] = gray[0][0]
angleGray[1] = gray[0][w-1]
angleGray[2] = gray[h-1][0]
angleGray[3] = gray[h-1][w-1]
num = 0
for i in range(len(angleGray)):
if angleGray[i] > 100:
num += 1
#print num
''' num大于3说明四个角是白色的，否则是黑色'''
if num >=3:
#把原来颜色反转后加强
ret , bin = cv2.threshold(gray, 100,255, cv2.THRESH_BINARY_INV)
else:
#保持原来颜色不变，只是加强
ret , bin = cv2.threshold(gray, 100,255, cv2.THRESH_BINARY)
#cv2.imshow("bin",bin)
#膨胀后腐蚀
dilated = cv2.dilate(bin, cv2.getStructuringElement(cv2.MORPH_RECT,(2, 2)))
eroded = cv2.erode(dilated, cv2.getStructuringElement(cv2.MORPH_RECT,(2, 2)))
# 腐蚀后膨胀
eroded = cv2.erode(eroded, cv2.getStructuringElement(cv2.MORPH_RECT,(2, 2)))
dilated = cv2.dilate(eroded, cv2.getStructuringElement(cv2.MORPH_RECT,(2, 2)))
#细化
median = cv2.medianBlur(dilated, 3)
median1 = cv2.medianBlur(dilated, 3)
#轮廓查找,查找前必须转换成黑底白字
contours, heirs = cv2.findContours(median1,cv2.RETR_EXTERNAL,cv2.CHAIN_APPROX_SIMPLE)
i = 0
pic = []
dictPic = {}
for tours in contours:
rc = cv2.boundingRect(tours)
#rc[0] 表示图像左上角的纵坐标，rc[1] 表示图像左上角的横坐标，rc[2] 表示图像的宽度，rc[3] 表示图像的高度，
#cv2.rectangle(bin, (rc[0],rc[1]),(rc[0]+rc[2],rc[1]+rc[3]),(255,0,255))
image1M = cv.fromarray(median)
image1Ip = cv.GetImage(image1M)
cv.SetImageROI(image1Ip,rc)
imageCopy = cv.CreateImage((rc[2], rc[3]),cv2.IPL_DEPTH_8U, 1)
cv.Copy(image1Ip,imageCopy)
cv.ResetImageROI(image1Ip)
#print np.asarray(cv.GetMat(imageCopy))
#把图像左上角的纵坐标和图像的数组元素放到字典里
dictPic[rc[0]] = np.asarray(cv.GetMat(imageCopy))
pic.append(np.asarray(cv.GetMat(imageCopy)))
#cv.ShowImage(str(i), imageCopy)
#cv.Not(imageCopy, imageCopy) #函数cvNot(const CvArr* src,CvArr* dst)会将src中的每一个元素的每一位取反，然后把结果赋给dst
#cv.SaveImage(str(i)+ '.jpg',imageCopy)
i = i+1
sortedNum = sorted(dictPic.keys())
for i in range(len(sortedNum)):
pic[i] = dictPic[sortedNum[i]]
#cv2.waitKey(0)
return pic
'''
#调整图片大小，先归一化图片，并把原图片放在中间,返回的数据格式 [[],[],.....]
'''
def resize(picArray,size):
picNew = []
for i in range(len(picArray)):
imgPIL = Image.fromarray(picArray[i])
h,w = imgPIL.size
newH = w//2 - h//2 #把图片放在中间
imgEmpty = Image.new('L',(w,w),0) #创建一张背景为黑色的图片
imgEmpty.paste(imgPIL,(newH,0))
imgResize = imgEmpty.resize(size,Image.ANTIALIAS)
imgResize0255 = imgResize.point(lambda x: 255 if x > 10 else 0) #0是黑，255是白 黑白加强
imgResizeArray = array(imgResize0255).flatten().tolist() #转换为一维
imgResizeArraySmaller = [float(x)/255 for x in imgResizeArray]#把0-255转成0-1
#print array(imgResize0255)
picNew.append(imgResizeArraySmaller)
#imgResize0255.show()
imgResize0255.save('D:/pic/result/' + str(i) + '.jpg')
return picNew
'''
#获取训练数据,path为文件夹路径,suffix为图片后缀，图片名字的首字母必须是图片中的内容
'''
def traindata(path,suffix):
train_images = []
train_labels = []
for files in glob.glob(path + '/*.' + suffix):
filepath,filename = os.path.split(files)
train_labels.append(int(filename[0:1]))
pic = picSplitResize(filepath + '/' + filename)
picNew = resize(pic,(20,20))
train_images.append(picNew[0])
#im = Image.open(filepath + '/' + filename)
return train_images,train_labels
'''
#创建LibSVM分类器,返回值为识别出的内容
'''
def predictPIC(train_images,train_labels,picdata):
prob = svm_problem(train_labels,train_images)
param = svm_parameter('-t 2')
prob = svm_train(prob,param)
labels = [0] * len(picdata)
flag = svm_predict(labels,picdata,prob)
print flag[0]
pic = picSplitResize('D:/pic/12.png')
#pic = picSplitResize('D:/pic/train and test/train/9-5.bmp')
#pic = picSplitResize('D:/pic/train and test/test1/7.bmp')
picdata = resize(pic,(20,20))
#train_images,train_labels = traindata('D:/pic/train and test/train','bmp')
#result = predictPIC(train_images,train_labels,picdata)
用于测试的图片
分割的结果(只贴出来部分)
cv2.imread()读入的图像使用NumPy数组表示，而通过cv.LoadImage()读入的图像为iplimage对象
array = cv2.imread("lena.jpg")
iplimage = cv.LoadImage("lena.jpg")
