教你使用TensorFlow2識別驗(yàn)證碼

更新時間：2021年06月10日 14:08:33 作者：劉潤森！

驗(yàn)證碼是根據(jù)隨機(jī)字符生成一幅圖片，然后在圖片中加入干擾象素，本文主要介紹了 TensorFlow2識別驗(yàn)證碼，需要的朋友們下面隨著小編來一起學(xué)習(xí)學(xué)習(xí)吧

驗(yàn)證碼是根據(jù)隨機(jī)字符生成一幅圖片，然后在圖片中加入干擾象素，用戶必須手動填入，防止有人利用機(jī)器人自動批量注冊、灌水、發(fā)垃圾廣告等等。

數(shù)據(jù)集來源：https://www.kaggle.com/fournierp/captcha-version-2-images

圖片是5個字母的單詞，可以包含數(shù)字。這些圖像應(yīng)用了噪聲（模糊和一條線）。它們是200 x 50 PNG。我們的任務(wù)是嘗試制作光學(xué)字符識別算法的模型。

在數(shù)據(jù)集中存在的驗(yàn)證碼png圖片，對應(yīng)的標(biāo)簽就是圖片的名字。

import os
import numpy as np
import pandas as pd
import cv2
import matplotlib.pyplot as plt
import seaborn as sns
# imgaug 圖片數(shù)據(jù)增強(qiáng)
import imgaug.augmenters as iaa
import tensorflow as tf
# Conv2D MaxPooling2D Dropout Flatten Dense BN  GAP
from tensorflow.keras.layers import Conv2D, MaxPooling2D, Dropout, Flatten, Dense, Layer, BatchNormalization, GlobalAveragePooling2D 
from tensorflow.keras.optimizers import Adam
from tensorflow.keras import Model, Input 
from tensorflow.keras.callbacks import EarlyStopping, ReduceLROnPlateau
# 圖片處理器
from tensorflow.keras.preprocessing.image import ImageDataGenerator
import plotly.express as px
import plotly.graph_objects as go
import plotly.offline as pyo
pyo.init_notebook_mode()

對數(shù)據(jù)進(jìn)行一個簡單的分析，統(tǒng)計(jì)圖像中大約出現(xiàn)了什么樣的符號。

# 數(shù)據(jù)路徑
DIR = '../input/captcha-version-2-images/samples/samples'
# 存儲驗(yàn)證碼的標(biāo)簽
captcha_list = []
characters = {}
for captcha in os.listdir(DIR):
    captcha_list.append(captcha)
    # 每張驗(yàn)證碼的captcha_code
    captcha_code = captcha.split(".")[0]
    for i in captcha_code:
        # 遍歷captcha_code 
        characters[i] = characters.get(i, 0) +1
symbols = list(characters.keys())
len_symbols = len(symbols)
print(f'圖像中只使用了{(lán)len_symbols}符號')

plt.bar(*zip(*characters.items()))
plt.title('Frequency of symbols')
plt.show()

如何提取圖像的數(shù)據(jù)建立X，y？？

# 如何提取圖像 建立 model  X 的shape  1070 * 50 * 200 * 1 
# y的shape 5 * 1070 * 19
 
for i, captcha in enumerate(captcha_list):
    captcha_code = captcha.split('.')[0]
    # cv2.IMREAD_GRAYSCALE 灰度圖
    captcha_cv2 = cv2.imread(os.path.join(DIR, captcha),cv2.IMREAD_GRAYSCALE)
    # 縮放
    captcha_cv2 = captcha_cv2 / 255.0
    # print(captcha_cv2.shape) (50, 200) 
    # 將captcha_cv2的(50, 200) 切換成(50, 200, 1)
    captcha_cv2 = np.reshape(captcha_cv2, img_shape)
    # (5,19)
    targs = np.zeros((len_captcha, len_symbols))
    
    for a, b in enumerate(captcha_code):
        targs[a, symbols.index(b)] = 1
    X[i] = captcha_cv2
    y[:, i] = targs

print("shape of X:", X.shape)
print("shape of y:", y.shape)

輸出如下

print("shape of X:", X.shape)
print("shape of y:", y.shape)

通過Numpy中random 隨機(jī)選擇數(shù)據(jù)，劃分訓(xùn)練集和測試集

# 生成隨機(jī)數(shù)
from numpy.random import default_rng

rng = default_rng(seed=1)
test_numbers = rng.choice(1070, size=int(1070*0.3), replace=False)
X_test = X[test_numbers]
X_full = np.delete(X, test_numbers,0)
y_test = y[:,test_numbers]
y_full = np.delete(y, test_numbers,1)

val_numbers = rng.choice(int(1070*0.7), size=int(1070*0.3), replace=False)

X_val = X_full[val_numbers]
X_train = np.delete(X_full, val_numbers,0)
y_val = y_full[:,val_numbers]
y_train = np.delete(y_full, val_numbers,1)

在此驗(yàn)證碼數(shù)據(jù)中，容易出現(xiàn)過擬合的現(xiàn)象，你可能會想到添加更多的新數(shù)據(jù)、添加正則項(xiàng)等，但這里使用數(shù)據(jù)增強(qiáng)的方法，特別是對于機(jī)器視覺的任務(wù)，數(shù)據(jù)增強(qiáng)技術(shù)尤為重要。

常用的數(shù)據(jù)增強(qiáng)操作：imgaug庫。imgaug是提供了各種圖像增強(qiáng)操作的python庫 https://github.com/aleju/imgaug。

imgaug幾乎包含了所有主流的數(shù)據(jù)增強(qiáng)的圖像處理操作，增強(qiáng)方法詳見github

# Sequential(C, R)	 尺寸增加了5倍，
# 選取一系列子增強(qiáng)器C作用于每張圖片的位置,第二個參數(shù)表示是否對每個batch的圖片應(yīng)用不同順序的Augmenter list     # rotate=(-8, 8)  旋轉(zhuǎn)
# iaa.CropAndPad  截取(crop)或者填充(pad),填充時,被填充區(qū)域?yàn)楹谏?
# px: 想要crop(negative values)的或者pad(positive values)的像素點(diǎn)。
# (top, right, bottom, left)
# 當(dāng)pad_mode=constant的時候選擇填充的值
aug =iaa.Sequential([iaa.CropAndPad(
    px=((0, 10), (0, 35), (0, 10), (0, 35)),
    pad_mode=['edge'],
    pad_cval=1
),iaa.Rotate(rotate=(-8,8))])

X_aug_train = None
y_aug_train = y_train
for i in range(40):
    X_aug = aug(images = X_train)
    if X_aug_train is not None:
        X_aug_train = np.concatenate([X_aug_train, X_aug], axis = 0)
        y_aug_train = np.concatenate([y_aug_train, y_train], axis = 1)
    else:
        X_aug_train = X_aug

讓我們看看一些數(shù)據(jù)增強(qiáng)的訓(xùn)練圖像。

fig, ax = plt.subplots(nrows=2, ncols =5, figsize = (16,16))
for i in range(10):
    index = np.random.randint(X_aug_train.shape[0])
    ax[i//5][i%5].imshow(X_aug_train[index],cmap='gray')

這次使用函數(shù)式API創(chuàng)建模型，函數(shù)式API是創(chuàng)建模型的另一種方式，它具有更多的靈活性，包括創(chuàng)建更為復(fù)雜的模型。

需要定義inputs和outputs

#函數(shù)式API模型創(chuàng)建
captcha = Input(shape=(50,200,channels))
x = Conv2D(32, (5,5),padding='valid',activation='relu')(captcha)
x = MaxPooling2D((2,2),padding='same')(x)
x = Conv2D(64, (3,3),padding='same',activation='relu')(x)
x = MaxPooling2D((2,2),padding='same')(x)
x = Conv2D(128, (3,3),padding='same',activation='relu')(x)
maxpool = MaxPooling2D((2,2),padding='same')(x)
outputs = []
for i in range(5):
    x = Conv2D(256, (3,3),padding='same',activation='relu')(maxpool)
    x = MaxPooling2D((2,2),padding='same')(x)
    x = Flatten()(x)
    x = Dropout(0.5)(x)
    x = BatchNormalization()(x)
    x = Dense(64, activation='relu')(x)
    x = Dropout(0.5)(x)
    x = BatchNormalization()(x)
    x = Dense(len_symbols , activation='softmax' , name=f'char_{i+1}')(x)
    outputs.append(x)
    
model = Model(inputs = captcha , outputs=outputs)
# ReduceLROnPlateau更新學(xué)習(xí)率
reduce_lr = ReduceLROnPlateau(patience =3, factor = 0.5,verbose = 1)
model.compile(loss='categorical_crossentropy', optimizer=Adam(learning_rate=0.0005), metrics=["accuracy"])
# EarlyStopping用于提前停止訓(xùn)練的callbacks。具體地,可以達(dá)到當(dāng)訓(xùn)練集上的loss不在減小
earlystopping = EarlyStopping(monitor ="val_loss",  
                             mode ="min", patience = 10,
                              min_delta = 1e-4,
                             restore_best_weights = True) 

history = model.fit(X_train, [y_train[i] for i in range(5)], batch_size=32, epochs=30, verbose=1, validation_data = (X_val, [y_val[i] for i in range(5)]), callbacks =[earlystopping,reduce_lr])

在這里插入圖片描述

下面對model進(jìn)行一個測試和評估。

score = model.evaluate(X_test,[y_test[0], y_test[1], y_test[2], y_test[3], y_test[4]],verbose=1)
metrics = ['loss','char_1_loss', 'char_2_loss', 'char_3_loss', 'char_4_loss', 'char_5_loss', 'char_1_acc', 'char_2_acc', 'char_3_acc', 'char_4_acc', 'char_5_acc']

for i,j in zip(metrics, score):
    print(f'{i}: {j}')

具體輸出如下：

11/11 [==============================] - 0s 11ms/step - loss: 0.7246 - char_1_loss: 0.0682 - char_2_loss: 0.1066 - char_3_loss: 0.2730 - char_4_loss: 0.2636 - char_5_loss: 0.0132 - char_1_accuracy: 0.9844 - char_2_accuracy: 0.9657 - char_3_accuracy: 0.9408 - char_4_accuracy: 0.9626 - char_5_accuracy: 0.9938
loss: 0.7246273756027222
char_1_loss: 0.06818050146102905
char_2_loss: 0.10664034634828568
char_3_loss: 0.27299806475639343
char_4_loss: 0.26359987258911133
char_5_loss: 0.013208594173192978
char_1_acc: 0.9844236969947815
char_2_acc: 0.9657320976257324
char_3_acc: 0.940809965133667
char_4_acc: 0.9626168012619019
char_5_acc: 0.9937694668769836

字母1到字母5的精確值都大于

繪制loss和score

metrics_df = pd.DataFrame(history.history)

columns = [col for col in metrics_df.columns if 'loss' in col and len(col)>8]

fig = px.line(metrics_df, y = columns)
fig.show()

plt.figure(figsize=(15,8))
plt.plot(history.history['loss'])
plt.plot(history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'val'], loc='upper right',prop={'size': 10})
plt.show()

# 預(yù)測數(shù)據(jù)
def predict(captcha):
    captcha = np.reshape(captcha , (1, 50,200,channels))
    result = model.predict(captcha)
    result = np.reshape(result ,(5,len_symbols))
    # 取出最大預(yù)測中的輸出
    label = ''.join([symbols[np.argmax(i)] for i in result])
    return label
    
predict(X_test[2])
# 25277

下面預(yù)測所有的數(shù)據(jù)

actual_pred = []

for i in range(X_test.shape[0]):
    actual = ''.join([symbols[i] for i in (np.argmax(y_test[:, i],axis=1))])
    pred =  predict(X_test[i])
    actual_pred.append((actual, pred))
print(actal_pred[:10])

輸出如下：

[('n4b4m', 'n4b4m'), ('42nxy', '42nxy'), ('25257', '25277'), ('cewnm', 'cewnm'), ('w46ep', 'w46ep'), ('cdcb3', 'edcb3'), ('8gf7n', '8gf7n'), ('nny5e', 'nny5e'), ('gm2c2', 'gm2c2'), ('g7fmc', 'g7fmc')]

sameCount = 0
diffCount = 0
letterDiff = {i:0 for i in range(5)}
incorrectness = {i:0 for i in range(1,6)}
for real, pred in actual_pred:
    # 預(yù)測和輸出相同
    if real == pred:
        sameCount += 1
    else:
        # 失敗
        diffCount += 1
        # 遍歷
        incorrectnessPoint = 0
        for i in range(5):
            if real[i] != pred[i]:
                letterDiff[i] += 1
                incorrectnessPoint += 1
        incorrectness[incorrectnessPoint] += 1


x = ['True predicted', 'False predicted']
y = [sameCount, diffCount]

fig = go.Figure(data=[go.Bar(x = x, y = y)])
fig.show()

在預(yù)測數(shù)據(jù)中，一共有287個數(shù)據(jù)預(yù)測正確。

在這里，我們可以看到出現(xiàn)錯誤到底是哪一個index。

x1 = ["Character " + str(x) for x in range(1, 6)]
    
fig = go.Figure(data=[go.Bar(x = x1, y = list(letterDiff.values()))])
fig.show()

為了計(jì)算每個單詞的錯誤數(shù)，繪制相關(guān)的條形圖。

x2 = [str(x) + " incorrect" for x in incorrectness.keys()]
y2 = list(incorrectness.values())

fig = go.Figure(data=[go.Bar(x = x2, y = y2)])
fig.show()

下面繪制錯誤的驗(yàn)證碼圖像，并標(biāo)準(zhǔn)正確和錯誤的區(qū)別。

fig, ax = plt.subplots(nrows = 8, ncols=4,figsize = (16,20))
count = 0
for i, (actual , pred) in enumerate(actual_pred):
    if actual != pred:
        img = X_test[i]
        try:
            ax[count//4][count%4].imshow(img, cmap = 'gray')
            ax[count//4][count%4].title.set_text(pred + ' - ' + actual)
            count += 1
        except:
            pass