OpenCV + svmutil 验证码识别

1 主程序

# -*- coding: utf-8 -*-
import cv2
import time
import numpy as np 
import pandas as pd




# 1 抓取2000张训练图片放在MIS文件夹，200张测试图片放在TEST文件夹
'''
url='http://mis.sgu.edu.cn/WaterMark.ashx'

for i in range(2000):
	urllib.request.urlretrieve(url,'01_MIS/%s.jpg'%i)
	if i%10==1:
		print(i)

for i in range(200):
	urllib.request.urlretrieve(url,'02_TEST/%s.jpg'%i)
	if i%10==1:
		print(i)





# 2 因为原始图片为gif格式，而opencv读取不了，
# 所以先用PIL把所有训练的图片转化为PNG格式，保存在03_TRAIN_PNG文件夹里面
from PIL import Image
for i in range(2000):
	try:
		im=Image.open('01_MIS/%s.jpg'%i)
		im.save('03_TRAIN_PNG/%s.png'%i)
		print(i)
	except:
		print('---------------------读取出错---------------------',i)
		continue




# 3 opencv图片进行预处理 放至文件夹 04_SLIPE
JPG=0
num=0
for i in range(2000):
	try:
		img = cv2.imread('03_TRAIN_PNG/%d.png'%i,0)
		# 图片进行二值化
		ret1,img = cv2.threshold(img,150,255,1)
		# 对验证码进行分割
		img1=img[4:,6:14]
		img2=img[4:,15:23]
		img3=img[4:,24:32]
		img4=img[4:,33:41]
		img5=img[4:,42:50]
		# 保存 后 去手工区分 0,1,2.。。。9
		cv2.imwrite('04_SLIPE/%d.png'%num,img1)
		num=num+1
		cv2.imwrite('04_SLIPE/%d.png'%num,img2)
		num=num+1
		cv2.imwrite('04_SLIPE/%d.png'%num,img3)
		num=num+1
		cv2.imwrite('04_SLIPE/%d.png'%num,img4)
		num=num+1
		cv2.imwrite('04_SLIPE/%d.png'%num,img5)
		num=num+1
		# 打印进度
		if i%10==1:
			print('slipe %s piture...'%i)

	except:
		print('--------------ERROR-------------',i,'num=',num)
		continue
# print(img1.shape,img2.shape,img3.shape,img4.shape,img5.shape)




# 5 手工把不同数字放进相应文件夹
# 1到9每个数字170个样本



# 6 读取图片转化为数字

H=0

txt_0=open('libsvm_train.txt','w')  # w 只写 r 只读
# 循环每个文件夹
for file in range(10):

	#循环读取每个文件夹下的图片，因为图片名字最高的图片2300多
	# 训练部分
	for i in range(2000):
		try:
			img = cv2.imread('05_NUM_0_9/%s/%s.png'%(file,i),0)#读取为灰度图片 3维读取为2维
			# 把最白的255 转为数值1
			img[img[:,:]>0] =1

			df_img=pd.DataFrame(img)

			txt_0.write(str(file))
			# print(df_img.sum())
			for col in range(8):
				txt_0.write(' %s:%s'%(col+1,  str(df_img.sum()[col])) )
			# print(df_img.sum(axis=1))
			for cum in range(14):
				txt_0.write(' %s:%s'%(cum+9,  str(df_img.sum(axis=1)[cum])) )
			txt_0.write('\n')

			H=H+1
			if H%20==1:
				print('piture %s '%H)
		except:
			continue

	print('正在处理数字 %s .....'%file)


txt_0.close()
# 读取
txt_0=open('libsvm_train.txt','r')
print(txt_0.read())


# 测试集
txt_2=open('libsvm_test.txt','w')  # w 只写 r 只读
# 循环每个文件夹
for file in range(10):

	#循环读取每个文件夹下的图片，因为图片名字最高的图片2300多
	# 训练部分
	for i in range(2000,2500):
		try:
			img = cv2.imread('05_NUM_0_9/%s/%s.png'%(file,i),0)#读取为灰度图片 3维读取为2维
			# 把最白的255 转为数值1
			img[img[:,:]>0] =1

			df_img=pd.DataFrame(img)

			txt_2.write(str(file))
			# print(df_img.sum())
			for col in range(8):
				txt_2.write(' %s:%s'%(col+1,  str(df_img.sum()[col])) )
			# print(df_img.sum(axis=1))
			for cum in range(14):
				txt_2.write(' %s:%s'%(cum+9,  str(df_img.sum(axis=1)[cum])) )
			txt_2.write('\n')

			H=H+1
			if H%20==1:
				print('piture %s '%H)
		except:
			continue

	print('正在处理数字 %s .....'%file)


txt_2.close()


'''

# SVM
from svmutil import *
'''
# 训练SVM模型
y,x=svm_read_problem('libsvm_train.txt')
model=svm_train(y,x,'-c 4')
# model=svm_train(y[:1200],x[:1200],'-c 4')
# 保存SVM模型
svm_save_model('libsvm_model.txt', model)
'''
# 读取SVM模型
model=svm_load_model('libsvm_model.txt')
test_y,test_x=svm_read_problem('libsvm_test.txt')
p_label,p_acc,p_val=svm_predict(test_y,test_x,model)

print(p_label)

2 读取模型，从文件夹里读取图片进行识别

# -*- coding: utf-8 -*-
import cv2
import time
import numpy as np 
import pandas as pd
from PIL import Image
import requests
from bs4 import BeautifulSoup
import urllib.request
from svmutil import *


model=svm_load_model('libsvm_model.txt')
url='http://mis.sgu.edu.cn/WaterMark.ashx'
for PIC in range(1000):

	svm_train_100=open('svm_train_100.txt','w')
	# 从文件读取识别
	img = cv2.imread('03_TRAIN_PNG/%d.png'%PIC,0)



	# 二值化
	ret1,img = cv2.threshold(img,150,255,1)
	# 把最白的255 转为数值1
	img[img[:,:]>0] =1
	# 切割图片
	img1=img[4:,6:14]
	img2=img[4:,15:23]
	img3=img[4:,24:32]
	img4=img[4:,33:41]
	img5=img[4:,42:50]
	# 读取为数值
	df_img1=pd.DataFrame(img1)
	df_img2=pd.DataFrame(img2)
	df_img3=pd.DataFrame(img3)
	df_img4=pd.DataFrame(img4)
	df_img5=pd.DataFrame(img5)

# 1
	svm_train_100.write('0')
	# print(df_img.sum())
	for col in range(8):
		svm_train_100.write(' %s:%s'%(col+1,  str(df_img1.sum()[col])) )
		# print(df_img.sum(axis=1))
	for cum in range(14):
		svm_train_100.write(' %s:%s'%(cum+9,  str(df_img1.sum(axis=1)[cum])) )
	svm_train_100.write('\n')


# 2
	svm_train_100.write('0')
	# print(df_img.sum())
	for col in range(8):
		svm_train_100.write(' %s:%s'%(col+1,  str(df_img2.sum()[col])) )
		# print(df_img.sum(axis=1))
	for cum in range(14):
		svm_train_100.write(' %s:%s'%(cum+9,  str(df_img2.sum(axis=1)[cum])) )
	svm_train_100.write('\n')


# 3
	svm_train_100.write('0')
	# print(df_img.sum())
	for col in range(8):
		svm_train_100.write(' %s:%s'%(col+1,  str(df_img3.sum()[col])) )
		# print(df_img.sum(axis=1))
	for cum in range(14):
		svm_train_100.write(' %s:%s'%(cum+9,  str(df_img3.sum(axis=1)[cum])) )
	svm_train_100.write('\n')

# 4
	svm_train_100.write('0')
	# print(df_img.sum())
	for col in range(8):
		svm_train_100.write(' %s:%s'%(col+1,  str(df_img4.sum()[col])) )
		# print(df_img.sum(axis=1))
	for cum in range(14):
		svm_train_100.write(' %s:%s'%(cum+9,  str(df_img4.sum(axis=1)[cum])) )
	svm_train_100.write('\n')


# 5
	svm_train_100.write('0')
	# print(df_img.sum())
	for col in range(8):
		svm_train_100.write(' %s:%s'%(col+1,  str(df_img5.sum()[col])) )
		# print(df_img.sum(axis=1))
	for cum in range(14):
		svm_train_100.write(' %s:%s'%(cum+9,  str(df_img5.sum(axis=1)[cum])) )
	svm_train_100.write('\n')

	svm_train_100.close()


	
	test_y,test_x=svm_read_problem('svm_train_100.txt')
	p_label,p_acc,p_val=svm_predict(test_y,test_x,model)
	print(p_label)
	NAME=''
	for i in range(5):
		NAME=NAME+str(int(p_label[i]))

	print(NAME)
	img = cv2.imread('03_TRAIN_PNG/%d.png'%PIC)
	cv2.imwrite('Result/%s.png'%NAME,img)