tf识别固定长度验证码图片ocr（0到9 4位）- CNN方式

我们先生成些验证码图片

import cv2 as cv
import numpy as np
import os


def create_digit_image(dir_path):
    image = np.ones(shape=[24, 72], dtype=np.uint8)
    image = image * 127
    a = np.random.randint(0, 10)
    b = np.random.randint(0, 10)
    c = np.random.randint(0, 10)
    d = np.random.randint(0, 10)
    text = str(a)+str(b)+str(c)+str(d)
    print(text)
    cv.putText(image, text, (6, 20), cv.FONT_HERSHEY_PLAIN, 1.5, (255), 2)
    for i in range(100):
        row = np.random.randint(0, 24)
        col = np.random.randint(0, 72)
        image[row, col] = 0
    full_path = dir_path + text + ".png"
    cv.imwrite(full_path, image)


os.mkdir(os.getcwd()+'\\train\\')
os.mkdir(os.getcwd()+'\\test\\')

for i in range(1000):
    create_digit_image(os.getcwd()+'\\train\\')

for i in range(100):
    create_digit_image(os.getcwd()+'\\test\\')　　

 

会生成1000张训练图片+100张测试图片

 

 

 

One-hot编码：

def text2vec(text):
    text_len = len(text)
    if text_len > 4:
        print("text code : ", text)
        raise ValueError('验证码最长4个字符')
    vector = np.zeros(4 * 10)

    def char2pos(c):
        k = ord(c)
        if 48 <= k <= 57:
            return k - 48
    for i, c in enumerate(text):
        idx = i * 10 + char2pos(c)
        vector[idx] = 1
    return vector


# 向量转回文本
def vec2text(vec):
    char_pos = vec.nonzero()[0]
    text = []
    for i, c in enumerate(char_pos):
        char_idx = c % 10
        if char_idx < 10:
            char_code = char_idx + ord('0')
        else:
            raise ValueError('error')
        text.append(chr(char_code))
    return "".join(text)



s=text2vec('1030')
print(s)

s = vec2text(s)
print(s)

　　

[0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
1030

　　

变化成如下：

[
  0. 1. 0. 0. 0. 0. 0. 0. 0. 0. 
  1. 0. 0. 0. 0. 0. 0. 0. 0. 0.
  0. 0. 0. 1. 0. 0. 0. 0. 0. 0.
  1. 0. 0. 0. 0. 0. 0. 0. 0. 0.
]

第一行代表1
第二行代表0
第三行代表3
第四行代表0

　　

完整代码：

import os
import tensorflow as tf
from random import choice
from tfdemo3.data_engine import get_one_image, get_image_files

w = 72
h = 24
label_vector_size = 40
train_dir = os.getcwd()+'\\train\\'
test_dir = os.getcwd()+'\\test\\'
train_files = get_image_files(train_dir)
test_files = get_image_files(test_dir)

# 占位符
x_image = tf.placeholder(shape=[None, h, w, 1], dtype=tf.float32)
y = tf.placeholder(shape=[None, label_vector_size], dtype=tf.float32)
keep_prob = tf.placeholder(dtype=tf.float32)

# convolution layer 1
conv1_w = tf.Variable(tf.random_normal(shape=[3, 3, 1, 32], stddev=0.1, dtype=tf.float32))
conv1_bias = tf.Variable(tf.random_normal(shape=[32], stddev=0.1))
conv1_out = tf.nn.conv2d(input=x_image, filter=conv1_w, strides=[1, 1, 1, 1], padding='SAME')
conv1_relu = tf.nn.relu(tf.add(conv1_out, conv1_bias))

# max pooling 1
maxpooling_1 = tf.nn.max_pool(conv1_relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

# convolution layer 2
conv2_w = tf.Variable(tf.random_normal(shape=[3, 3, 32, 64], stddev=0.1, dtype=tf.float32))
conv2_bias = tf.Variable(tf.random_normal(shape=[64], stddev=0.1))
conv2_out = tf.nn.conv2d(input=maxpooling_1, filter=conv2_w, strides=[1, 1, 1, 1], padding='SAME')
conv2_relu = tf.nn.relu(tf.add(conv2_out, conv2_bias))

# max pooling 2
maxpooling_2 = tf.nn.max_pool(conv2_relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

# convolution layer 3
conv3_w = tf.Variable(tf.random_normal(shape=[3, 3, 64, 64], stddev=0.1, dtype=tf.float32))
conv3_bias = tf.Variable(tf.random_normal(shape=[64], stddev=0.1))
conv3_out = tf.nn.conv2d(input=maxpooling_2, filter=conv3_w, strides=[1, 1, 1, 1], padding='SAME')
conv3_relu = tf.nn.relu(tf.add(conv3_out, conv3_bias))

# max pooling 2
maxpooling_3 = tf.nn.max_pool(conv3_relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

# fc-1
w_fc1 = tf.Variable(tf.random_normal(shape=[3*9*64, 1024], stddev=0.1, dtype=tf.float32))
b_fc1 = tf.Variable(tf.constant(0.1, shape=[1024]))
h_pool2 = tf.reshape(maxpooling_3, [-1, 3*9*64])
output_fc1 = tf.nn.relu(tf.add(tf.matmul(h_pool2, w_fc1), b_fc1))

# dropout
h2 = tf.nn.dropout(output_fc1, keep_prob=keep_prob)

# fc-2
w_fc2 = tf.Variable(tf.random_normal(shape=[1024, 40], stddev=0.1, dtype=tf.float32))
b_fc2 = tf.Variable(tf.constant(0.1, shape=[40]))
y_conv = tf.add(tf.matmul(output_fc1, w_fc2), b_fc2)

# loss
cross_loss = tf.nn.sigmoid_cross_entropy_with_logits(logits=y_conv, labels=y)
loss = tf.reduce_mean(cross_loss)
step = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)

# accuracy
saver = tf.train.Saver()
predict = tf.reshape(y_conv, [-1, 4, 10])
max_idx_p = tf.argmax(predict, 2)
max_idx_l = tf.argmax(tf.reshape(y, [-1, 4, 10]), 2)
correct_pred = tf.equal(max_idx_p, max_idx_l)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))


def get_train_batch(files, batch_size=128):
    images = []
    labels = []
    for f in range(batch_size):
        image, label = get_one_image(train_dir, choice(files))
        images.append(image)
        labels.append(label)
    return images, labels


def get_batch(root_dir, files):
    images = []
    labels = []
    for f in files:
        image, label = get_one_image(root_dir, f)
        images.append(image)
        labels.append(label)
    return images, labels


test_images, test_labels = get_batch(test_dir, test_files)


with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(500):
        batch_xs, batch_ys = get_train_batch(train_files, 100)
        curr_loss, curr_ = sess.run([loss, step], feed_dict={x_image: batch_xs, y: batch_ys, keep_prob: 0.5})
        if (i + 1) % 100 == 0:
            print("run step (%d) ..., loss : (%f)" % (i+1, curr_loss))
            curr_acc = sess.run(accuracy, feed_dict={x_image: test_images, y: test_labels, keep_prob: 1.0})
            print("current test Accuracy : %f" % (curr_acc))
    saver.save(sess, "./ckp/code_break.ckpt", global_step=500)

　　

data_engine.py

import numpy as np
import cv2 as cv
import os


def text2vec(text):
    text_len = len(text)
    if text_len > 4:
        print("text code : ", text)
        raise ValueError('验证码最长4个字符')
    vector = np.zeros(4 * 10)

    def char2pos(c):
        k = ord(c)
        if 48 <= k <= 57:
            return k - 48
    for i, c in enumerate(text):
        idx = i * 10 + char2pos(c)
        vector[idx] = 1
    return vector


# 向量转回文本
def vec2text(vec):
    char_pos = vec.nonzero()[0]
    text = []
    for i, c in enumerate(char_pos):
        char_idx = c % 10
        if char_idx < 10:
            char_code = char_idx + ord('0')
        else:
            raise ValueError('error')
        text.append(chr(char_code))
    return "".join(text)


def get_one_image(root_dir, f):
    gray = cv.imread(os.path.join(root_dir, f), cv.IMREAD_GRAYSCALE)
    resize = cv.resize(gray, (72, 24))
    result = np.zeros(resize.shape, dtype=np.float32)
    cv.normalize(resize, result, 0, 1, cv.NORM_MINMAX, dtype=cv.CV_32F)
    image = np.expand_dims(result, axis=2)
    label = text2vec(f[0:4])
    return image, label


def get_image_files(root_dir):
    img_list = []
    files = os.listdir(root_dir)
    for f in files:
        if os.path.isfile(os.path.join(root_dir, f)):
            img_list.append(f)
    return img_list

　　

 

run step (100) ..., loss : (0.023609)
current test Accuracy : 0.992500
run step (200) ..., loss : (0.000665)
current test Accuracy : 1.000000
run step (300) ..., loss : (0.000046)
current test Accuracy : 1.000000
run step (400) ..., loss : (0.000010)
current test Accuracy : 1.000000
run step (500) ..., loss : (0.000005)
current test Accuracy : 1.000000

　　

卷积网络确实比较好。