文章目录
- 1）引入相关模块
- 2）设置变量参数
- 3）函数定义
- 生成4位验证码函数
- 验证码标签和图片生成函数
- 图像转换
- 文字转化函数
- 向量转化函数
- 生成训练数据
- 4）搭建神经网络
- 搭建网络结构
- 配置训练参数
- 执行训练过程
- 5）训练结果
代码转载于:
将模型的训练过程和生成数据做了简化，搭建卷积神经网络对10000张灰色图片进行识别，在迭代到40次时，准确率可达到100%
1）引入相关模块
from captcha.image import ImageCaptcha
import random
from PIL import Image
import numpy as np
import tensorflow as tf模块介绍：
Python中图像处理库PIL的Image模块，用Image.open(+地址)打开图像。
Python中验证码生成库captcha
from captcha.image import ImageCaptcha
chars = '14gh'# 随意选取4个字母，看验证码生成效果
image = ImageCaptcha().generate_image(chars)
image.show()
image.save("test.png")
2）设置变量参数
其中
SAVE_PATH所设置的是模型的保存地址
CHAR_SET得到了由26个英文字母（大写+小写，共52个）加上10个数字组成的一个大的列表
CHAR_SET_LEN值为上述列表的长度，为62
IMAGE_HEIGHT为需要识别的验证码图片的高度为60
IMAGE_WIDTH为需要识别的验证码图片的宽度为160
number = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']
alphabet = ['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u',
'v', 'w', 'x', 'y', 'z']
ALPHABET = ['A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M', 'N', 'O', 'P', 'Q', 'R', 'S', 'T', 'U',
'V', 'W', 'X', 'Y', 'Z']
SAVE_PATH = "D:/tensorflow_demo/deep learning/"
CHAR_SET = number + alphabet + ALPHABET # 得到的CHAR_SET是一个包含62个元素的大列表
CHAR_SET_LEN = len(CHAR_SET) # 26个小写字母+26个大写字母+10个数字 共62个
IMAGE_HEIGHT = 60
IMAGE_WIDTH = 1603）函数定义
生成4位验证码函数
# 该函数的作用是随机生成4位验证码 返回的captcha_text是一个含有四位字母或数字组成的列表
def random_captcha_text(char_set=None, captcha_size=4):
if char_set is None:
char_set = number + alphabet + ALPHABET
captcha_text = []
for i in range(captcha_size):
c = random.choice(char_set)
captcha_text.append(c)
return captcha_text验证码标签和图片生成函数
def gen_captcha_text_and_image(width=160, height=60, char_set=CHAR_SET):
image = ImageCaptcha(width=width, height=height)
captcha_text = random_captcha_text(char_set) # 调用已定义好的random_captcha_text函数 得到了随机的四位验证码的列表
captcha_text = ''.join(captcha_text)
captcha = image.generate(captcha_text) # 调用验证码生成函数，captcha为生成的验证码图片
captcha_image = Image.open(captcha)
captcha_image = np.array(captcha_image) # 将图片用数组像素表示
return captcha_text, captcha_image
text, image = gen_captcha_text_and_image(char_set=CHAR_SET)
MAX_CAPTCHA = len(text)
print('CHAR_SET_LEN=', CHAR_SET_LEN, ' MAX_CAPTCHA=', MAX_CAPTCHA)图像转换
验证码的识别与是否为彩色照片无关，所以在此将其转化为灰度图像，可以减少训练的参数
该函数传入的为生成的验证码图片
def convert2gray(img):
if len(img.shape) > 2:
gray = np.mean(img, -1)
return gray
else:
return img文字转化函数
该函数传入的是验证码的标签文字内容，返回的是一个vector，4*62维度的，每一行对应文字为1，其余为0
def text2vec(text):
vector = np.zeros([MAX_CAPTCHA, CHAR_SET_LEN]) # max_captcha为4，char_set_len为62
for i, c in enumerate(text): # enumerate是枚举的含义 i是索引从0到4，c为4个验证码文字
idx = CHAR_SET.index(c) # idx为4位验证码文字在62个大列表中的索引
vector[i][idx] = 1.0
return vectorPS:补充enumerate函数的用法
enumerate是python的内置函数，利用它可以获得索引和值
1）、对于一个列表，若想同时获得索引和值，可以这样编写
list1 = ['hello','world','I','love','python']
for i in range(len(list1)):
print(i,list1[i])
0 hello
1 world
2 I
3 love
4 python编写起来较为麻烦
2）、使用enumerate函数会更加方便
list1 = ['hello','world','I','love','python']
for index,value in enumerate(list1):
print(index,value)
0 hello
1 world
2 I
3 love
4 pythonenumerate函数还可以接收第二个参数，就是索引开始的初始值
list1 = ['hello','world','I','love','python']
for index,value in enumerate(list1,1):
print(index,value)
1 hello
2 world
3 I
4 love
5 python向量转化函数
输入的vec是一个4*62的矩阵，返回的text为四位验证码文字
def vec2text(vec):
text = []
for i, c in enumerate(vec):
text.append(CHAR_SET[c])
return "".join(text)生成训练数据
获得了10000张验证码的灰色图片和标签
x_train = np.zeros([10000, IMAGE_HEIGHT, IMAGE_WIDTH, 1]) # batch_x为生成的验证码图片，通道数为1的灰色图片
y_train = np.zeros([10000, MAX_CAPTCHA, CHAR_SET_LEN]) # batch_y为标签，4*62的矩阵
for i in range(10000):
text,image = gen_captcha_text_and_image(char_set=CHAR_SET)
image = tf.reshape(convert2gray(image), (IMAGE_HEIGHT, IMAGE_WIDTH, 1)) # （60，160，1）
x_train[i, :] = image
y_train[i, :] = text2vec(text)4）搭建神经网络
搭建网络结构
model = tf.keras.Sequential()
model.add(tf.keras.layers.Conv2D(32, (3, 3)))
model.add(tf.keras.layers.PReLU())
model.add(tf.keras.layers.MaxPool2D((2, 2), strides=2))
model.add(tf.keras.layers.Conv2D(64, (5, 5)))
model.add(tf.keras.layers.PReLU())
model.add(tf.keras.layers.MaxPool2D((2, 2), strides=2))
model.add(tf.keras.layers.Conv2D(128, (5, 5)))
model.add(tf.keras.layers.PReLU())
model.add(tf.keras.layers.MaxPool2D((2, 2), strides=2))
model.add(tf.keras.layers.Flatten())
model.add(tf.keras.layers.Dense(MAX_CAPTCHA * CHAR_SET_LEN)) # Dense全连接的输出为4*62
model.add(tf.keras.layers.Reshape([MAX_CAPTCHA, CHAR_SET_LEN]))
model.add(tf.keras.layers.Softmax())配置训练参数
model.compile(optimizer='Adam',
metrics=['accuracy'],
loss='categorical_crossentropy')执行训练过程
history = model.fit(x_train, y_train, batch_size=32, epochs=50)5）训练结果
Epoch 1/50
313/313 [==============================] - 202s 644ms/step - loss: 4.6302 - accuracy: 0.0144
Epoch 2/50
313/313 [==============================] - 204s 650ms/step - loss: 4.4329 - accuracy: 0.0154
Epoch 3/50
313/313 [==============================] - 204s 653ms/step - loss: 4.1740 - accuracy: 0.0159
Epoch 4/50
313/313 [==============================] - 213s 681ms/step - loss: 14.4912 - accuracy: 0.0166
Epoch 5/50
313/313 [==============================] - 201s 643ms/step - loss: 7.9571 - accuracy: 0.0157
Epoch 6/50
313/313 [==============================] - 158s 505ms/step - loss: 4.1492 - accuracy: 0.0167
Epoch 7/50
313/313 [==============================] - 159s 507ms/step - loss: 4.1425 - accuracy: 0.0168
Epoch 8/50
313/313 [==============================] - 165s 528ms/step - loss: 4.1385 - accuracy: 0.0175
Epoch 9/50
313/313 [==============================] - 170s 542ms/step - loss: 4.1365 - accuracy: 0.0169
Epoch 10/50
313/313 [==============================] - 169s 539ms/step - loss: 4.1319 - accuracy: 0.0196
Epoch 11/50
313/313 [==============================] - 179s 570ms/step - loss: 4.0545 - accuracy: 0.0336
Epoch 12/50
313/313 [==============================] - 172s 549ms/step - loss: 3.5705 - accuracy: 0.1045
Epoch 13/50
313/313 [==============================] - 173s 552ms/step - loss: 3.0129 - accuracy: 0.1980
Epoch 14/50
313/313 [==============================] - 171s 545ms/step - loss: 2.6300 - accuracy: 0.2831
Epoch 15/50
313/313 [==============================] - 170s 542ms/step - loss: 2.2148 - accuracy: 0.3930
Epoch 16/50
313/313 [==============================] - 181s 580ms/step - loss: 1.7232 - accuracy: 0.5208
Epoch 17/50
313/313 [==============================] - 186s 593ms/step - loss: 1.2240 - accuracy: 0.6564
Epoch 18/50
313/313 [==============================] - 179s 573ms/step - loss: 0.8103 - accuracy: 0.7691
Epoch 19/50
313/313 [==============================] - 179s 572ms/step - loss: 0.4705 - accuracy: 0.8638
Epoch 20/50
313/313 [==============================] - 195s 624ms/step - loss: 0.2919 - accuracy: 0.9146
Epoch 21/50
313/313 [==============================] - 189s 605ms/step - loss: 0.2021 - accuracy: 0.9373
Epoch 22/50
313/313 [==============================] - 179s 571ms/step - loss: 0.1775 - accuracy: 0.9453
Epoch 23/50
313/313 [==============================] - 189s 604ms/step - loss: 0.1705 - accuracy: 0.9459
Epoch 24/50
313/313 [==============================] - 188s 600ms/step - loss: 0.1571 - accuracy: 0.9515
Epoch 25/50
313/313 [==============================] - 162s 517ms/step - loss: 0.1666 - accuracy: 0.9505
Epoch 26/50
313/313 [==============================] - 162s 517ms/step - loss: 0.1300 - accuracy: 0.9607
Epoch 27/50
313/313 [==============================] - 163s 520ms/step - loss: 0.1494 - accuracy: 0.9560
Epoch 28/50
313/313 [==============================] - 162s 517ms/step - loss: 0.1653 - accuracy: 0.9508
Epoch 29/50
313/313 [==============================] - 159s 507ms/step - loss: 0.1281 - accuracy: 0.9635
Epoch 30/50
313/313 [==============================] - 160s 510ms/step - loss: 0.1336 - accuracy: 0.9619
Epoch 31/50
313/313 [==============================] - 159s 507ms/step - loss: 0.1323 - accuracy: 0.9633
Epoch 32/50
313/313 [==============================] - 159s 507ms/step - loss: 0.1346 - accuracy: 0.9621
Epoch 33/50
313/313 [==============================] - 169s 540ms/step - loss: 0.1369 - accuracy: 0.9639
Epoch 34/50
313/313 [==============================] - 169s 541ms/step - loss: 0.1353 - accuracy: 0.9644
Epoch 35/50
313/313 [==============================] - 170s 543ms/step - loss: 0.1585 - accuracy: 0.9599
Epoch 36/50
313/313 [==============================] - 170s 543ms/step - loss: 4.2418 - accuracy: 0.7325
Epoch 37/50
313/313 [==============================] - 170s 542ms/step - loss: 0.1098 - accuracy: 0.9720
Epoch 38/50
313/313 [==============================] - 169s 539ms/step - loss: 0.0211 - accuracy: 0.9938
Epoch 39/50
313/313 [==============================] - 169s 540ms/step - loss: 0.0063 - accuracy: 0.9983
Epoch 40/50
313/313 [==============================] - 169s 540ms/step - loss: 0.0025 - accuracy: 0.9994
Epoch 41/50
313/313 [==============================] - 170s 543ms/step - loss: 4.4409e-04 - accuracy: 0.9999
Epoch 42/50
313/313 [==============================] - 168s 538ms/step - loss: 1.6461e-04 - accuracy: 1.0000
Epoch 43/50
313/313 [==============================] - 169s 540ms/step - loss: 9.7764e-05 - accuracy: 1.0000
Epoch 44/50
313/313 [==============================] - 168s 538ms/step - loss: 7.7123e-05 - accuracy: 1.0000
Epoch 45/50
313/313 [==============================] - 169s 541ms/step - loss: 6.3981e-05 - accuracy: 1.0000
Epoch 46/50
313/313 [==============================] - 169s 540ms/step - loss: 5.4254e-05 - accuracy: 1.0000
Epoch 47/50
313/313 [==============================] - 169s 541ms/step - loss: 4.6260e-05 - accuracy: 1.0000
Epoch 48/50
313/313 [==============================] - 168s 537ms/step - loss: 3.9702e-05 - accuracy: 1.0000
Epoch 49/50
313/313 [==============================] - 168s 538ms/step - loss: 3.4279e-05 - accuracy: 1.0000
Epoch 50/50
313/313 [==============================] - 168s 537ms/step - loss: 2.9718e-05 - accuracy: 1.0000PS：补充模型的保存与调用
首先需要指定模型保存的路径
SAVE_PATH = "D:/tensorflow_demo/deep learning/"模型的保存
try:
model = tf.keras.models.load_model(SAVE_PATH + 'model') # 将模型保存到指定路径下
except Exception as e:
print('#######Exception', e)当再次需要调用时
model = tf.keras.models.load_model(SAVE_PATH + 'model')
本文章为转载内容，我们尊重原作者对文章享有的著作权。如有内容错误或侵权问题，欢迎原作者联系我们进行内容更正或删除文章。
