TensorFlow驗證碼識別

本節咱們來用 TensorFlow 來實現一個深度學習模型,用來實現驗證碼識別的過程,這裏咱們識別的驗證碼是圖形驗證碼,首先咱們會用標註好的數據來訓練一個模型,而後再用模型來實現這個驗證碼的識別。python

驗證碼

首先咱們來看下驗證碼是怎樣的,這裏咱們使用 Python 的 captcha 庫來生成便可,這個庫默認是沒有安裝的,因此這裏咱們須要先安裝這個庫,另外咱們還須要安裝 pillow 庫,使用 pip3 便可:git

pip3 install captcha pillow

安裝好以後,咱們就能夠用以下代碼來生成一個簡單的圖形驗證碼了:github

from captcha.image import ImageCaptchafrom PIL import Imagetext = '1234'image = ImageCaptcha()captcha = image.generate(text)captcha_image = Image.open(captcha)captcha_image.show()

運行以後便會彈出一張圖片,結果以下:api

圖片

能夠看到圖中的文字正是咱們所定義的 text 內容,這樣咱們就能夠獲得一張圖片和其對應的真實文本,這樣咱們就能夠用它來生成一批訓練數據和測試數據了。數組

預處理

在訓練以前確定是要進行數據預處理了,如今咱們首先定義好了要生成的驗證碼文本內容,這就至關於已經有了 label 了,而後咱們再用它來生成驗證碼,就能夠獲得輸入數據 x 了,在這裏咱們首先定義好咱們的輸入詞表,因爲大小寫字母加數字的詞表比較龐大,設想咱們用含有大小寫字母和數字的驗證碼,一個驗證碼四個字符,那麼一共可能的組合是 (26 + 26 + 10) ^ 4 = 14776336 種組合,這個數量訓練起來有點大,因此這裏咱們精簡一下,只使用純數字的驗證碼來訓練,這樣其組合個數就變爲 10 ^ 4 = 10000 種,顯然少了不少。網絡

因此在這裏咱們先定義一個詞表和其長度變量:app

VOCAB = ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']CAPTCHA_LENGTH = 4VOCAB_LENGTH = len(VOCAB)

這裏 VOCAB 就是詞表的內容,即 0 到 9 這 10 個數字,驗證碼的字符個數即 CAPTCHA_LENGTH 是 4,詞表長度是 VOCAB 的長度,即 10。dom

接下來咱們定義一個生成驗證碼數據的方法,流程相似上文,只不過這裏咱們將返回的數據轉爲了 Numpy 形式的數組:ide

from PIL import Imagefrom captcha.image import ImageCaptchaimport numpy as npdef generate_captcha(captcha_text):    """    get captcha text and np array    :param captcha_text: source text    :return: captcha image and array    """    image = ImageCaptcha()    captcha = image.generate(captcha_text)    captcha_image = Image.open(captcha)    captcha_array = np.array(captcha_image)    return captcha_array

這樣調用此方法,咱們就能夠獲得一個 Numpy 數組了,這個實際上是把驗證碼轉化成了每一個像素的 RGB,咱們調用一下這個方法試試:函數

captcha = generate_captcha('1234')print(captcha, captcha.shape)

內容以下:

[[[239 244 244]  [239 244 244]  [239 244 244]  ...,   ...,   [239 244 244]  [239 244 244]  [239 244 244]]] (60, 160, 3)

能夠看到它的 shape 是 (60, 160, 3),這其實表明驗證碼圖片的高度是 60,寬度是 160,是 60 x 160 像素的驗證碼,每一個像素都有 RGB 值,因此最後一維即爲像素的 RGB 值。

接下來咱們須要定義 label,因爲咱們須要使用深度學習模型進行訓練,因此這裏咱們的 label 數據最好使用 One-Hot 編碼,即若是驗證碼文本是 1234,那麼應該詞表索引位置置 1,總共的長度是 40,咱們用程序實現一下 One-Hot 編碼和文本的互相轉換:

def text2vec(text):    """    text to one-hot vector    :param text: source text    :return: np array    """    if len(text) > CAPTCHA_LENGTH:        return False    vector = np.zeros(CAPTCHA_LENGTH * VOCAB_LENGTH)    for i, c in enumerate(text):        index = i * VOCAB_LENGTH + VOCAB.index(c)        vector[index] = 1    return vectordef vec2text(vector):    """    vector to captcha text    :param vector: np array    :return: text    """    if not isinstance(vector, np.ndarray):        vector = np.asarray(vector)    vector = np.reshape(vector, [CAPTCHA_LENGTH, -1])    text = ''    for item in vector:        text += VOCAB[np.argmax(item)]    return text

這裏 text2vec() 方法就是將真實文本轉化爲 One-Hot 編碼,vec2text() 方法就是將 One-Hot 編碼轉回真實文本。

例如這裏調用一下這兩個方法,咱們將 1234 文本轉換爲 One-Hot 編碼,而後在將其轉回來:

vector = text2vec('1234')text = vec2text(vector)print(vector, text)

運行結果以下:

[ 0.  1.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  1.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  1.  0.  0.  0.  0.  0.  0.  0.  0.  0.  0.  1.  0.  0.  0.  0.  0.]1234

這樣咱們就能夠實現文本到 One-Hot 編碼的互轉了。

接下來咱們就能夠構造一批數據了,x 數據就是驗證碼的 Numpy 數組,y 數據就是驗證碼的文本的 One-Hot 編碼,生成內容以下:

import randomfrom os.path import join, existsimport pickleimport numpy as npfrom os import makedirsDATA_LENGTH = 10000DATA_PATH = 'data'def get_random_text():    text = ''    for i in range(CAPTCHA_LENGTH):        text += random.choice(VOCAB)    return textdef generate_data():    print('Generating Data...')    data_x, data_y = [], []    # generate data x and y    for i in range(DATA_LENGTH):        text = get_random_text()        # get captcha array        captcha_array = generate_captcha(text)        # get vector        vector = text2vec(text)        data_x.append(captcha_array)        data_y.append(vector)    # write data to pickle    if not exists(DATA_PATH):        makedirs(DATA_PATH)    x = np.asarray(data_x, np.float32)    y = np.asarray(data_y, np.float32)    with open(join(DATA_PATH, 'data.pkl'), 'wb') as f:        pickle.dump(x, f)        pickle.dump(y, f)

這裏咱們定義了一個 getrandomtext() 方法,能夠隨機生成驗證碼文本,而後接下來再利用這個隨機生成的文原本產生對應的 x、y 數據,而後咱們再將數據寫入到 pickle 文件裏,這樣就完成了預處理的操做。

構建模型

有了數據以後,咱們就開始構建模型吧,這裏咱們仍是利用 traintestsplit() 方法將數據分爲三部分,訓練集、開發集、驗證集:

with open('data.pkl', 'rb') as f:    data_x = pickle.load(f)    data_y = pickle.load(f)    return standardize(data_x), data_ytrain_x, test_x, train_y, test_y = train_test_split(data_x, data_y, test_size=0.4, random_state=40)dev_x, test_x, dev_y, test_y, = train_test_split(test_x, test_y, test_size=0.5, random_state=40)

接下來咱們使用者三個數據集構建三個 Dataset 對象:

# train and dev datasettrain_dataset = tf.data.Dataset.from_tensor_slices((train_x, train_y)).shuffle(10000)train_dataset = train_dataset.batch(FLAGS.train_batch_size)dev_dataset = tf.data.Dataset.from_tensor_slices((dev_x, dev_y))dev_dataset = dev_dataset.batch(FLAGS.dev_batch_size)test_dataset = tf.data.Dataset.from_tensor_slices((test_x, test_y))test_dataset = test_dataset.batch(FLAGS.test_batch_size)

而後初始化一個迭代器,並綁定到這個數據集上:

# a reinitializable iteratoriterator = tf.data.Iterator.from_structure(train_dataset.output_types, train_dataset.output_shapes)train_initializer = iterator.make_initializer(train_dataset)dev_initializer = iterator.make_initializer(dev_dataset)test_initializer = iterator.make_initializer(test_dataset)

接下來就是關鍵的部分了,在這裏咱們使用三層卷積和兩層全鏈接網絡進行構造,在這裏爲了簡化寫法,直接使用 TensorFlow 的 layers 模塊:

# input Layerwith tf.variable_scope('inputs'):    # x.shape = [-1, 60, 160, 3]    x, y_label = iterator.get_next()keep_prob = tf.placeholder(tf.float32, [])y = tf.cast(x, tf.float32)# 3 CNN layersfor _ in range(3):    y = tf.layers.conv2d(y, filters=32, kernel_size=3, padding='same', activation=tf.nn.relu)    y = tf.layers.max_pooling2d(y, pool_size=2, strides=2, padding='same')    # y = tf.layers.dropout(y, rate=keep_prob)# 2 dense layersy = tf.layers.flatten(y)y = tf.layers.dense(y, 1024, activation=tf.nn.relu)y = tf.layers.dropout(y, rate=keep_prob)y = tf.layers.dense(y, VOCAB_LENGTH)

這裏卷積核大小爲 3,padding 使用 SAME 模式,激活函數使用 relu。

通過全鏈接網絡變換以後,y 的 shape 就變成了 [batchsize, nclasses],咱們的 label 是 CAPTCHALENGTH 個 One-Hot 向量拼合而成的,在這裏咱們想使用交叉熵來計算,可是交叉熵計算的時候,label 參數向量最後一維各個元素之和必須爲 1,否則計算梯度的時候會出現問題。詳情參見 TensorFlow 的官方文檔:https://www.tensorflow.org/apidocs/python/tf/nn/softmaxcrossentropywithlogits:

NOTE: While the classes are mutually exclusive, their probabilities need not be. All that is required is that each row of labels is a valid probability distribution. If they are not, the computation of the gradient will be incorrect.

可是如今的 label 參數是 CAPTCHALENGTH 個 One-Hot 向量拼合而成,因此這裏各個元素之和爲 CAPTCHALENGTH,因此咱們須要從新 reshape 一下,確保最後一維各個元素之和爲 1:

y_reshape = tf.reshape(y, [-1, VOCAB_LENGTH])y_label_reshape = tf.reshape(y_label, [-1, VOCAB_LENGTH])

這樣咱們就能夠確保最後一維是 VOCAB_LENGTH 長度,而它就是一個 One-Hot 向量,因此各元素之和一定爲 1。

而後 Loss 和 Accuracy 就好計算了:

# losscross_entropy = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(logits=y_reshape, labels=y_label_reshape))# accuracymax_index_predict = tf.argmax(y_reshape, axis=-1)max_index_label = tf.argmax(y_label_reshape, axis=-1)correct_predict = tf.equal(max_index_predict, max_index_label)accuracy = tf.reduce_mean(tf.cast(correct_predict, tf.float32))

再接下來執行訓練便可:

# traintrain_op = tf.train.RMSPropOptimizer(FLAGS.learning_rate).minimize(cross_entropy, global_step=global_step)for epoch in range(FLAGS.epoch_num):    tf.train.global_step(sess, global_step_tensor=global_step)    # train    sess.run(train_initializer)    for step in range(int(train_steps)):        loss, acc, gstep, _ = sess.run([cross_entropy, accuracy, global_step, train_op],                                       feed_dict={keep_prob: FLAGS.keep_prob})        # print log        if step % FLAGS.steps_per_print == 0:            print('Global Step', gstep, 'Step', step, 'Train Loss', loss, 'Accuracy', acc)    if epoch % FLAGS.epochs_per_dev == 0:        # dev        sess.run(dev_initializer)        for step in range(int(dev_steps)):            if step % FLAGS.steps_per_print == 0:                print('Dev Accuracy', sess.run(accuracy, feed_dict={keep_prob: 1}), 'Step', step)

在這裏咱們首先初始化 traininitializer,將 iterator 綁定到 Train Dataset 上,而後執行 trainop,得到 loss、acc、gstep 等結果並輸出。

訓練

運行訓練過程,結果相似以下:

...Dev Accuracy 0.9580078 Step 0Dev Accuracy 0.9472656 Step 2Dev Accuracy 0.9501953 Step 4Dev Accuracy 0.9658203 Step 6Global Step 3243 Step 0 Train Loss 1.1920928e-06 Accuracy 1.0Global Step 3245 Step 2 Train Loss 1.5497207e-06 Accuracy 1.0Global Step 3247 Step 4 Train Loss 1.1920928e-06 Accuracy 1.0Global Step 3249 Step 6 Train Loss 1.7881392e-06 Accuracy 1.0...

驗證集準確率 95% 以上。

測試

訓練過程咱們還能夠每隔幾個 Epoch 保存一下模型:

# save modelif epoch % FLAGS.epochs_per_save == 0:    saver.save(sess, FLAGS.checkpoint_dir, global_step=gstep)

固然也能夠取驗證集上準確率最高的模型進行保存。

驗證時咱們能夠從新 Reload 一下模型,而後進行驗證:

# load modelckpt = tf.train.get_checkpoint_state('ckpt')if ckpt:    saver.restore(sess, ckpt.model_checkpoint_path)    print('Restore from', ckpt.model_checkpoint_path)    sess.run(test_initializer)    for step in range(int(test_steps)):        if step % FLAGS.steps_per_print == 0:            print('Test Accuracy', sess.run(accuracy, feed_dict={keep_prob: 1}), 'Step', step)else:    print('No Model Found')

驗證以後其準確率基本是差很少的。

若是要進行新的 Inference 的話,能夠替換下 test_x 便可。

代碼

以上即是使用 TensorFlow 進行驗證碼識別的過程,代碼見:https://github.com/AIDeepLearning/CrackCaptcha。

相關文章
相關標籤/搜索