RNN入門(二)識別驗證碼
阿新 • • 發佈:2018-12-12
介紹
作為RNN的第二個demo,筆者將會介紹RNN模型在識別驗證碼方面的應用。 我們的驗證碼及樣本資料集來自於部落格: CNN大戰驗證碼,在這篇部落格中,我們已經準備好了所需的樣本資料集,不需要在辛辛苦苦地再弄一遍,直接呼叫data.csv就可以進行建模了。
RNN模型
用TensorFlow搭建簡單RNN模型,因為是多分類問題,所以在最後的輸出部分再加一softmax層,損失函式採用對數損失函式,optimizer選擇RMSPropOptimizer。以下是RNN模型的完整Python程式碼(TensorFlow_RNN.py):
# -*- coding: utf-8 -*-
import tensorflow as tf
import logging
# 設定日誌
logging.basicConfig(level = logging.INFO, format='%(asctime)s - %(levelname)s: %(message)s')
logger = logging.getLogger(__name__)
# RNN類
class RNN:
# 初始化
# 引數說明: element_size: 元素大小
# time_steps: 序列大小
# num_classes: 目標變數的類別總數
# batch_size: 圖片總數
# hidden_layer_size: 隱藏層的神經元個數
# epoch: 訓練次數
# learning_rate: 用RMSProp優化時的學習率
# save_model_path: 模型儲存地址
def __init__(self, element_size, time_steps, num_classes, batch_size, hidden_layer_size = 150,
epoch = 1000, learning_rate=0.001, save_model_path = r'./logs/RNN_train.ckpt'):
self.epoch = epoch
self.learning_rate = learning_rate
self.save_model_path = save_model_path
# 設定RNN結構
self.element_size = element_size
self.time_steps = time_steps
self.num_classes = num_classes
self.batch_size = batch_size
self.hidden_layer_size = hidden_layer_size
# 輸入向量和輸出向量
self._inputs = tf.placeholder(tf.float32, shape=[None, self.time_steps, self.element_size], name='inputs')
self.y = tf.placeholder(tf.float32, shape=[None, self.num_classes], name='inputs')
# 利用TensorFlow的內建函式BasicRNNCell, dynamic_rnn來構建RNN的基本模組
rnn_cell = tf.contrib.rnn.BasicRNNCell(self.hidden_layer_size)
outputs, _ = tf.nn.dynamic_rnn(rnn_cell, self._inputs, dtype=tf.float32)
Wl = tf.Variable(tf.truncated_normal([self.hidden_layer_size, self.num_classes], mean=0, stddev=.01))
bl = tf.Variable(tf.truncated_normal([self.num_classes], mean=0, stddev=.01))
def get_linear_layer(vector):
return tf.matmul(vector, Wl) + bl
# 取輸出的向量outputs中的最後一個向量最為最終輸出
last_rnn_output = outputs[:, -1, :]
self.final_output = get_linear_layer(last_rnn_output)
# 定義損失函式並用RMSProp優化
softmax = tf.nn.softmax_cross_entropy_with_logits(logits=self.final_output, labels=self.y)
self.cross_entropy = tf.reduce_mean(softmax)
self.train_model = tf.train.RMSPropOptimizer(self.learning_rate, 0.9).minimize(self.cross_entropy)
self.saver = tf.train.Saver()
logger.info('Initialize RNN model...')
# 模型訓練
def train(self, x_data, y_data):
logger.info('Training RNN model...')
with tf.Session() as sess:
# 對所有變數進行初始化
sess.run(tf.global_variables_initializer())
# 進行迭代學習
feed_dict = {self._inputs: x_data, self.y: y_data}
for i in range(self.epoch + 1):
sess.run(self.train_model, feed_dict=feed_dict)
if i % int(self.epoch / 50) == 0:
# to see the step improvement
print('已訓練%d次, loss: %s.' % (i, sess.run(self.cross_entropy, feed_dict=feed_dict)))
# 儲存RNN模型
logger.info('Saving RNN model...')
self.saver.save(sess, self.save_model_path)
# 對新資料進行預測
def predict(self, data):
with tf.Session() as sess:
logger.info('Restoring RNN model...')
self.saver.restore(sess, self.save_model_path)
predict = sess.run(self.final_output, feed_dict={self._inputs: data})
return predict
模型訓練
對樣本資料集data.csv進行RNN建模,將資料集分為訓練集和測試集,各佔70%和30%.因為圖片的大小為16*20,所以在將圖片看成序列時,序列的長度為20,每一時刻的向量含有16個元素,共有31個目標類,取隱藏層大小為300,總共訓練1000次。 完整的Python程式碼如下:
# -*- coding: utf-8 -*-
"""
數字字母識別
利用RNN對驗證碼的資料集進行多分類
"""
from TensorFlow_RNN import RNN
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score
from sklearn.preprocessing import LabelBinarizer
CSV_FILE_PATH = 'F://驗證碼識別/data.csv' # CSV 檔案路徑
df = pd.read_csv(CSV_FILE_PATH) # 讀取CSV檔案
# 資料集的特徵
features = ['v'+str(i+1) for i in range(16*20)]
labels = df['label'].unique()
# 對樣本的真實標籤進行標籤二值化
lb = LabelBinarizer()
lb.fit(labels)
y_ture = pd.DataFrame(lb.transform(df['label']), columns=['y'+str(i) for i in range(31)])
y_bin_columns = list(y_ture.columns)
for col in y_bin_columns:
df[col] = y_ture[col]
# 將資料集分為訓練集和測試集,訓練集70%, 測試集30%
x_train, x_test, y_train, y_test = train_test_split(df[features], df[y_bin_columns], \
train_size = 0.7, test_size=0.3, random_state=123)
# 構建RNN網路
# 模型儲存地址
MODEL_SAVE_PATH = 'F:///驗證碼識別/logs/RNN_train.ckpt'
# RNN初始化
element_size = 16
time_steps = 20
num_classes = 31
hidden_layer_size = 300
batch_size = 960
new_x_train = np.array(x_train).reshape((-1, time_steps, element_size))
new_x_test = np.array(x_test).reshape((-1, time_steps, element_size))
rnn = RNN(element_size=element_size,
time_steps=time_steps,
num_classes=num_classes,
batch_size=batch_size,
hidden_layer_size= hidden_layer_size,
epoch=1000,
save_model_path=MODEL_SAVE_PATH,
)
# 訓練RNN
rnn.train(new_x_train, y_train)
# 預測資料
y_pred = rnn.predict(new_x_test)
# 預測分類
label = '123456789ABCDEFGHJKLNPQRSTUVXYZ'
prediction = []
for pred in y_pred:
label = labels[list(pred).index(max(pred))]
prediction.append(label)
# 計算預測的準確率
x_test['prediction'] = prediction
x_test['label'] = df['label'][y_test.index]
print(x_test.head())
accuracy = accuracy_score(x_test['prediction'], x_test['label'])
print('CNN的預測準確率為%.2f%%.'%(accuracy*100))
以下是模型訓練的結果:
2018-09-26 11:18:12,339 - INFO: Initialize RNN model...
2018-09-26 11:18:12,340 - INFO: Training RNN model...
已訓練0次, loss: 3.43417.
已訓練20次, loss: 3.42695.
已訓練40次, loss: 3.40638.
已訓練60次, loss: 3.33286.
已訓練80次, loss: 2.78305.
已訓練100次, loss: 2.33391.
已訓練120次, loss: 1.15871.
已訓練140次, loss: 0.659932.
已訓練160次, loss: 0.566225.
已訓練180次, loss: 0.397372.
已訓練200次, loss: 0.317218.
已訓練220次, loss: 0.346782.
已訓練240次, loss: 0.639625.
已訓練260次, loss: 0.0575929.
已訓練280次, loss: 0.100429.
已訓練300次, loss: 0.0347529.
已訓練320次, loss: 0.0189503.
已訓練340次, loss: 0.0265893.
已訓練360次, loss: 0.0151181.
已訓練380次, loss: 1.18662.
已訓練400次, loss: 0.0164317.
已訓練420次, loss: 0.00819814.
已訓練440次, loss: 0.0041992.
已訓練460次, loss: 0.0206414.
已訓練480次, loss: 0.00826409.
已訓練500次, loss: 0.00398952.
已訓練520次, loss: 0.00214751.
已訓練540次, loss: 0.0365587.
已訓練560次, loss: 0.00738376.
已訓練580次, loss: 0.00302118.
已訓練600次, loss: 0.00161713.
已訓練620次, loss: 0.000885372.
已訓練640次, loss: 1.24874.
已訓練660次, loss: 0.00601175.
已訓練680次, loss: 0.0023275.
已訓練700次, loss: 0.00121995.
已訓練720次, loss: 0.000705643.
已訓練740次, loss: 0.000407971.
已訓練760次, loss: 0.000219642.
已訓練780次, loss: 0.0889083.
已訓練800次, loss: 0.00395974.
已訓練820次, loss: 0.00131215.
已訓練840次, loss: 0.000631665.
已訓練860次, loss: 0.000342329.
已訓練880次, loss: 0.000191806.
已訓練900次, loss: 0.000108547.
已訓練920次, loss: 6.29806e-05.
已訓練940次, loss: 3.99281e-05.
已訓練960次, loss: 0.0124334.
已訓練980次, loss: 0.00142853.
2018-09-26 11:26:08,302 - INFO: Saving RNN model...
已訓練1000次, loss: 0.000571731.
2018-09-26 11:26:08,761 - INFO: Restoring RNN model...
INFO:tensorflow:Restoring parameters from F:///驗證碼識別/logs/RNN_train.ckpt
2018-09-26 11:26:08,761 - INFO: Restoring parameters from F:///驗證碼識別/logs/RNN_train.ckpt
v1 v2 v3 v4 v5 v6 v7 v8 v9 v10 ... v313 v314 v315 v316 \
657 1 1 1 1 1 1 1 1 1 1 ... 1 1 1 1
18 1 1 1 1 1 1 1 1 1 1 ... 1 1 1 1
700 1 1 1 1 1 1 1 1 1 1 ... 1 1 1 1
221 1 1 1 1 1 1 1 1 1 1 ... 1 1 1 1
1219 1 1 1 1 1 1 1 1 1 1 ... 1 1 1 1
v317 v318 v319 v320 prediction label
657 1 1 1 1 G G
18 1 1 1 1 1 1
700 1 1 1 1 H H
221 1 1 1 1 5 5
1219 1 1 1 1 V V
[5 rows x 322 columns]
CNN的預測準確率為93.69%.
總共的訓練時間為8分鐘,在測試集上的準確為93.69%.與CNN相比,測試集上的準確率略高,訓練時間卻明顯減少,因為CNN訓練1000次的時間為75分鐘。總的來說,該RNN模型在這個資料集的表現優於之前的CNN模型。
模型預測
接著,我們利用剛才訓練好的CNN模型,對新驗證碼進行識別,看看模型的識別效果。 筆者採集了50張新驗證碼,如下:
完整的預測新驗證碼的Python指令碼如下:
# -*- coding: utf-8 -*-
"""
利用訓練好的RNN模型對驗證碼進行識別
(共訓練960條資料,訓練1000次測試集上的準確率為95.15%.)
"""
import os
import cv2
import pandas as pd
import numpy as np
from TensorFlow_RNN import RNN
def split_picture(imagepath):
# 以灰度模式讀取圖片
gray = cv2.imread(imagepath, 0)
# 將圖片的邊緣變為白色
height, width = gray.shape
for i in range(width):
gray[0, i] = 255
gray[height-1, i] = 255
for j in range(height):
gray[j, 0] = 255
gray[j, width-1] = 255
# 中值濾波
blur = cv2.medianBlur(gray, 3) #模板大小3*3
# 二值化
ret,thresh1 = cv2.threshold(blur, 200, 255, cv2.THRESH_BINARY)
# 提取單個字元
chars_list = []
image, contours, hierarchy = cv2.findContours(thresh1, 2, 2)
for cnt in contours:
# 最小的外接矩形
x, y, w, h = cv2.boundingRect(cnt)
if x != 0 and y != 0 and w*h >= 100:
chars_list.append((x,y,w,h))
sorted_chars_list = sorted(chars_list, key=lambda x:x[0])
for i,item in enumerate(sorted_chars_list):
x, y, w, h = item
cv2.imwrite('F://chars/%d.jpg'%(i+1), thresh1[y:y+h, x:x+w])
def remove_edge_picture(imagepath):
image = cv2.imread(imagepath, 0)
height, width = image.shape
corner_list = [image[0,0] < 127,
image[height-1, 0] < 127,
image[0, width-1]<127,
image[ height-1, width-1] < 127
]
if sum(corner_list) >= 3:
os.remove(imagepath)
def resplit_with_parts(imagepath, parts):
image = cv2.imread(imagepath, 0)
os.remove(imagepath)
height, width = image.shape
file_name = imagepath.split('/')[-1].split(r'.')[0]
# 將圖片重新分裂成parts部分
step = width//parts # 步長
start = 0 # 起始位置
for i in range(parts):
cv2.imwrite('F://chars/%s.jpg'%(file_name+'-'+str(i)), \
image[:, start:start+step])
start += step
def resplit(imagepath):
image = cv2.imread(imagepath, 0)
height, width = image.shape
if width >= 64:
resplit_with_parts(imagepath, 4)
elif width >= 48:
resplit_with_parts(imagepath, 3)
elif width >= 26:
resplit_with_parts(imagepath, 2)
# rename and convert to 16*20 size
def convert(dir, file):
imagepath = dir+'/'+file
# 讀取圖片
image = cv2.imread(imagepath, 0)
# 二值化
ret, thresh = cv2.threshold(image, 127, 255, cv2.THRESH_BINARY)
img = cv2.resize(thresh, (16, 20), interpolation=cv2.INTER_AREA)
# 儲存圖片
cv2.imwrite('%s/%s' % (dir, file), img)
# 讀取圖片的資料,並轉化為0-1值
def Read_Data(dir, file):
imagepath = dir+'/'+file
# 讀取圖片
image = cv2.imread(imagepath, 0)
# 二值化
ret, thresh = cv2.threshold(image, 127, 255, cv2.THRESH_BINARY)
# 顯示圖片
bin_values = [1 if pixel==255 else 0 for pixel in thresh.ravel()]
return bin_values
def predict(rnn, VerifyCodePath, time_steps, element_size):
dir = 'F://chars'
files = os.listdir(dir)
# 清空原有的檔案
if files:
for file in files:
os.remove(dir + '/' + file)
split_picture(VerifyCodePath)
files = os.listdir(dir)
if not files:
print('檢視的資料夾為空!')
else:
# 去除噪聲圖片
for file in files:
remove_edge_picture(dir + '/' + file)
# 對黏連圖片進行重分割
for file in os.listdir(dir):
resplit(dir + '/' + file)
# 將圖片統一調整至16*20大小
for file in os.listdir(dir):
convert(dir, file)
# 圖片中的字元代表的向量
files = sorted(os.listdir(dir), key=lambda x: x[0])
table = [Read_Data(dir, file) for file in files]
test_data = pd.DataFrame(table, columns=['v%d' % i for i in range(1, 321)])
new_test_data = np.array(test_data).reshape((-1, time_steps, element_size))
y_pred = rnn.predict(new_test_data)
# 預測分類
prediction = []
labels = '123456789ABCDEFGHJKLNPQRSTUVXYZ'
for pred in y_pred:
label = labels[list(pred).index(max(pred))]
prediction.append(label)
TRUE_LABEL = VerifyCodePath.split('/')[-1].split(r'.')[0]
return TRUE_LABEL, ''.join(prediction)
def main():
# 建立RNN預測模型
# 模型儲存地址
MODEL_SAVE_PATH = 'F:///驗證碼識別/logs/RNN_train.ckpt'
# RNN初始化
element_size = 16
time_steps = 20
num_classes = 31
batch_size = 4
hidden_layer_size = 300
rnn = RNN(element_size=element_size,
time_steps=time_steps,
num_classes=num_classes,
batch_size=batch_size,
hidden_layer_size=hidden_layer_size,
epoch=1000,
save_model_path=MODEL_SAVE_PATH,
)
# 預測驗證碼
pred_list = []
dir = 'F://VerifyCode/'
for file in os.listdir(dir):
VerifyCodePath = dir+file
label, prediction = predict(rnn, VerifyCodePath, time_steps, element_size)
pred_list.append((label, prediction))
# print('真實值為:%s, 預測結果為: %s.'%(label, prediction))
# 統計預測正確的驗證碼的數量及準確率
total_images = len(pred_list)
correct_pred