1、知識點
""" 驗證碼分析： 對圖片進行分析： 1、分割識別 2、整體識別 輸出：[3,5,7] -->softmax轉為概率[0.04,0.16,0.8] ---> 交叉熵計算損失值 (目標值和預測值的對數) tf.argmax(預測值，2)
驗證碼樣例：[NAZP] [XCVB] [WEFW] ，都是字母的 """
2、將數據寫入TFRecords
import tensorflow as tf import os os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2' FLAGS = tf.app.flags.FLAGS tf.app.flags.DEFINE_string("tfrecords_dir", "./tfrecords/captcha.tfrecords", "驗證碼tfrecords文件") tf.app.flags.DEFINE_string("captcha_dir", "../data/Genpics/", "驗證碼圖片路徑") tf.app.flags.DEFINE_string("letter", "ABCDEFGHIJKLMNOPQRSTUVWXYZ", "驗證碼字符的種類") def dealwithlabel(label_str): # 構建字符索引 {0：'A', 1:'B'......} num_letter = dict(enumerate(list(FLAGS.letter))) # 鍵值對反轉 {'A':0, 'B':1......} letter_num = dict(zip(num_letter.values(), num_letter.keys())) print(letter_num) # 構建標簽的列表 array = [] # 給標簽數據進行處理[[b"NZPP"]......] for string in label_str: letter_list = []# [1,2,3,4] # 修改編碼，b'FVQJ'到字符串，並且循環找到每張驗證碼的字符對應的數字標記 for letter in string.decode('utf-8'): letter_list.append(letter_num[letter]) array.append(letter_list) # [[13, 25, 15, 15], [22, 10, 7, 10], [22, 15, 18, 9], [16, 6, 13, 10], [1, 0, 8, 17], [0, 9, 24, 14].....] print(array) # 將array轉換成tensor類型 label = tf.constant(array) return label def get_captcha_image(): """ 獲取驗證碼圖片數據 :param file_list: 路徑+文件名列表 :return: image """ # 構造文件名 filename = [] for i in range(6000): string = str(i) + ".jpg" filename.append(string) # 構造路徑+文件 file_list = [os.path.join(FLAGS.captcha_dir, file) for file in filename] # 構造文件隊列 file_queue = tf.train.string_input_producer(file_list, shuffle=False) # 構造閱讀器 reader = tf.WholeFileReader() # 讀取圖片數據內容 key, value = reader.read(file_queue) # 解碼圖片數據 image = tf.image.decode_jpeg(value) image.set_shape([20, 80, 3]) # 批處理數據 [6000, 20, 80, 3] image_batch = tf.train.batch([image], batch_size=6000, num_threads=1, capacity=6000) return image_batch def get_captcha_label(): """ 讀取驗證碼圖片標簽數據 :return: label """ file_queue = tf.train.string_input_producer(["../data/Genpics/labels.csv"], shuffle=False) reader = tf.TextLineReader() key, value = reader.read(file_queue) records = [[1], ["None"]] number, label = tf.decode_csv(value, record_defaults=records) # [["NZPP"], ["WKHK"], ["ASDY"]] label_batch = tf.train.batch([label], batch_size=6000, num_threads=1, capacity=6000) return label_batch def write_to_tfrecords(image_batch, label_batch): """ 將圖片內容和標簽寫入到tfrecords文件當中 :param image_batch: 特征值 :param label_batch: 標簽紙 :return: None """ # 轉換類型 label_batch = tf.cast(label_batch, tf.uint8) print(label_batch) # 建立TFRecords 存儲器 writer = tf.python_io.TFRecordWriter(FLAGS.tfrecords_dir) # 循環將每一個圖片上的數據構造example協議塊，序列化后寫入 for i in range(6000): # 取出第i個圖片數據，轉換相應類型,圖片的特征值要轉換成字符串形式 image_string = image_batch[i].eval().tostring() # 標簽值，轉換成整型 label_string = label_batch[i].eval().tostring() # 構造協議塊 example = tf.train.Example(features=tf.train.Features(feature={ "image": tf.train.Feature(bytes_list=tf.train.BytesList(value=[image_string])), "label": tf.train.Feature(bytes_list=tf.train.BytesList(value=[label_string])) })) writer.write(example.SerializeToString()) # 關閉文件 writer.close() return None if __name__ == "__main__": # 獲取驗證碼文件當中的圖片 image_batch = get_captcha_image() # 獲取驗證碼文件當中的標簽數據 label = get_captcha_label() print(image_batch, label) with tf.Session() as sess: coord = tf.train.Coordinator() threads = tf.train.start_queue_runners(sess=sess, coord=coord) # [b'NZPP' b'WKHK' b'WPSJ' ..., b'FVQJ' b'BQYA' b'BCHR'] label_str = sess.run(label) print(label_str) # 處理字符串標簽到數字張量 label_batch = dealwithlabel(label_str) print(label_batch) # 將圖片數據和內容寫入到tfrecords文件當中 write_to_tfrecords(image_batch, label_batch) coord.request_stop() coord.join(threads)
3、數據存在百度雲(小白號)
4、標准代碼
import tensorflow as tf FLAGS = tf.app.flags.FLAGS tf.app.flags.DEFINE_string("captcha_dir", "./tfrecords/captcha.tfrecords", "驗證碼數據的路徑") tf.app.flags.DEFINE_integer("batch_size", 100, "每批次訓練的樣本數") tf.app.flags.DEFINE_integer("label_num", 4, "每個樣本的目標值數量") tf.app.flags.DEFINE_integer("letter_num", 26, "每個目標值取的字母的可能心個數") # 定義一個初始化權重的函數 def weight_variables(shape): w = tf.Variable(tf.random_normal(shape=shape, mean=0.0, stddev=1.0)) return w # 定義一個初始化偏置的函數 def bias_variables(shape): b = tf.Variable(tf.constant(0.0, shape=shape)) return b def read_and_decode(): """ 讀取驗證碼數據API :return: image_batch, label_batch """ # 1、構建文件隊列 file_queue = tf.train.string_input_producer([FLAGS.captcha_dir]) # 2、構建閱讀器，讀取文件內容，默認一個樣本 reader = tf.TFRecordReader() # 讀取內容 key, value = reader.read(file_queue) # tfrecords格式example,需要解析 features = tf.parse_single_example(value, features={ "image": tf.FixedLenFeature([], tf.string), "label": tf.FixedLenFeature([], tf.string), }) # 解碼內容，字符串內容 # 1、先解析圖片的特征值 image = tf.decode_raw(features["image"], tf.uint8) # 1、先解析圖片的目標值 label = tf.decode_raw(features["label"], tf.uint8) # print(image, label) # 改變形狀 image_reshape = tf.reshape(image, [20, 80, 3]) label_reshape = tf.reshape(label, [4]) print(image_reshape, label_reshape) # 進行批處理,每批次讀取的樣本數 100, 也就是每次訓練時候的樣本 image_batch, label_btach = tf.train.batch([image_reshape, label_reshape], batch_size=FLAGS.batch_size, num_threads=1, capacity=FLAGS.batch_size) print(image_batch, label_btach) return image_batch, label_btach def fc_model(image): """ 進行預測結果 :param image: 100圖片特征值[100, 20, 80, 3] :return: y_predict預測值[100, 4 * 26] """ with tf.variable_scope("model"): # 將圖片數據形狀轉換成二維的形狀 image_reshape = tf.reshape(image, [-1, 20 * 80 * 3]) # 1、隨機初始化權重偏置 # matrix[100, 20 * 80 * 3] * [20 * 80 * 3, 4 * 26] + [104] = [100, 4 * 26] weights = weight_variables([20 * 80 * 3, 4 * 26]) bias = bias_variables([4 * 26]) # 進行全連接層計算[100, 4 * 26] y_predict = tf.matmul(tf.cast(image_reshape, tf.float32), weights) + bias return y_predict def predict_to_onehot(label): """ 將讀取文件當中的目標值轉換成one-hot編碼 :param label: [100, 4] [[13, 25, 15, 15], [19, 23, 20, 16]......] :return: one-hot """ # 進行one_hot編碼轉換，提供給交叉熵損失計算，准確率計算[100, 4, 26] label_onehot = tf.one_hot(label, depth=FLAGS.letter_num, on_value=1.0, axis=2) print(label_onehot) return label_onehot def captcharec(): """ 驗證碼識別程序 :return: """ # 1、讀取驗證碼的數據文件 label_btch [100 ,4] image_batch, label_batch = read_and_decode() # 2、通過輸入圖片特征數據，建立模型，得出預測結果 # 一層，全連接神經網絡進行預測 # matrix [100, 20 * 80 * 3] * [20 * 80 * 3, 4 * 26] + [104] = [100, 4 * 26] y_predict = fc_model(image_batch) # [100, 4 * 26] print(y_predict) # 3、先把目標值轉換成one-hot編碼 [100, 4, 26] y_true = predict_to_onehot(label_batch) # 4、softmax計算, 交叉熵損失計算 with tf.variable_scope("soft_cross"): # 求平均交叉熵損失 ,y_true [100, 4, 26]--->[100, 4*26] loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits( labels=tf.reshape(y_true, [FLAGS.batch_size, FLAGS.label_num * FLAGS.letter_num]), logits=y_predict)) # 5、梯度下降優化損失 with tf.variable_scope("optimizer"): train_op = tf.train.GradientDescentOptimizer(0.01).minimize(loss) # 6、求出樣本的每批次預測的准確率是多少 三維比較 with tf.variable_scope("acc"): # 比較每個預測值和目標值是否位置(4)一樣 y_predict [100, 4 * 26]---->[100, 4, 26] equal_list = tf.equal(tf.argmax(y_true, 2), tf.argmax(tf.reshape(y_predict, [FLAGS.batch_size, FLAGS.label_num, FLAGS.letter_num]), 2)) # equal_list 100個樣本 [1, 0, 1, 0, 1, 1,..........] accuracy = tf.reduce_mean(tf.cast(equal_list, tf.float32)) # 定義一個初始化變量的op init_op = tf.global_variables_initializer() # 開啟會話訓練 with tf.Session() as sess: sess.run(init_op) # 定義線程協調器和開啟線程（有數據在文件當中讀取提供給模型） coord = tf.train.Coordinator() # 開啟線程去運行讀取文件操作 threads = tf.train.start_queue_runners(sess, coord=coord) # 訓練識別程序 for i in range(5000): sess.run(train_op) print("第%d批次的准確率為：%f" % (i, accuracy.eval())) # 回收線程 coord.request_stop() coord.join(threads) return None if __name__ == "__main__": captcharec()
5、自寫代碼
# coding = utf-8 import tensorflow as tf from tensorflow.contrib.slim.python.slim.nets.inception_v3 import inception_v3_base import os """ 驗證碼分析： 對圖片進行分析： 1、分割識別 2、整體識別 輸出：[3,5,7] -->softmax轉為概率[0.04,0.16,0.8] ---> 交叉熵計算損失值 (目標值和預測值的對數) tf.argmax(預測值，2) """ FLAGS = tf.app.flags.FLAGS tf.app.flags.DEFINE_string("captcha_dir","./tfrecords/captcha.tfrecords","驗證碼數據路徑") tf.app.flags.DEFINE_integer("batch_size",100,"讀取批次") tf.app.flags.DEFINE_integer("label_num", 4, "每個樣本的目標值數量") tf.app.flags.DEFINE_integer("letter_num", 26, "每個目標值取的字母的可能心個數") def weight_variable(shape): w = tf.Variable(tf.random_normal(shape=shape,mean=0.0,stddev=1.0,)) return w def bias_variable(shape): b = tf.Variable(tf.random_normal(shape=shape,mean=0.0,stddev=1.0)) return b def captcharec(): """ 驗證碼識別 :return: """ #1、讀取驗證碼的數據文件 file_queue = tf.train.string_input_producer([FLAGS.captcha_dir]) #2、創建閱讀器,解析example reader = tf.TFRecordReader() key ,value = reader.read(file_queue) features = tf.parse_single_example(value,features={ "image":tf.FixedLenFeature([],tf.string), "label": tf.FixedLenFeature([], tf.string) }) #解碼操作 image = tf.decode_raw(features["image"],tf.uint8) label = tf.decode_raw(features["label"],tf.uint8) print(image,label) #修改形狀 image_reshape = tf.reshape(image,[20,80,3]) label_reshape = tf.reshape(label, [4]) #進行批處理，每次讀取100個樣本 image_batch,label_batch = tf.train.batch([image_reshape,label_reshape],batch_size=100,num_threads=1,capacity=20) print(image_batch, label_batch) return image_batch,label_batch def fc_model(image_batch): #1、初始化權重和偏置 w = weight_variable([20*80*3,4*26]) b = bias_variable([4*26]) #模型 x [100,20*80*3] w [20*80*3,4] y_true [100,4] #對輸入進行矩陣轉換 image = tf.reshape(image_batch,[-1,20*80*3]) y_predict = tf.matmul(tf.cast(image,tf.float32),w) + b ############收集變量######## tf.summary.histogram("w",w) tf.summary.histogram("b",b) merged = tf.summary.merge_all() return y_predict,merged #[100,4] def predict_to_onehot(label_batch): y_true = tf.one_hot(label_batch,on_value=1.0,depth=26,axis=2) return y_true if __name__ == '__main__': image_batch, label_batch = captcharec() y_predict,merged_his =fc_model(image_batch) y_true = predict_to_onehot(label_batch) #計算交叉熵 loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=tf.reshape(y_true,[100,4*26]),logits=y_predict)) #梯度下降優化 train_op = tf.train.GradientDescentOptimizer(0.001).minimize(loss) #准確率 equal_list = tf.equal(tf.argmax(y_true,2),tf.argmax(tf.reshape(y_predict,[100,4,26]),2)) accuracy = tf.reduce_mean(tf.cast(equal_list,tf.float32)) #####收集變量############### tf.summary.scalar("losses",loss) tf.summary.scalar("accuracy",accuracy) merged_scalar = tf.summary.merge_all() ############保存模型#### saver = tf.train.Saver() init_op = tf.global_variables_initializer() with tf.Session() as sess: sess.run(init_op) fileWriter = tf.summary.FileWriter("./vc",graph=sess.graph) #創建線程協調器 coord = tf.train.Coordinator() #開啟線程 threads = tf.train.start_queue_runners(sess,coord=coord) IS_TRAIN =1 # if os.path.exists("./vertifycode/checkpoint"): # IS_TRAIN = 0 if IS_TRAIN==1: #######訓練模型############### # if os.path.exists("./vertifycode/checkpoint"): # saver.restore(sess, "./vertifycode/vertifycode_model") for i in range(2000): sess.run(train_op) summary_his = sess.run(merged_his) summary_scalar = sess.run(merged_scalar) fileWriter.add_summary(summary_scalar,i) fileWriter.add_summary(summary_his,i) print("訓練第%d次的准確率為：%f" %(i,accuracy.eval())) #######保存模型############# saver.save(sess,"./vertifycode/vertifycode_model") else: ##########測試模型################## for i in range(10): saver.restore(sess, "./vertifycode/vertifycode_model") # print("第%d張圖片的准確率為：%f" % ( # i, # tf.argmax(y_test, 2).eval(), # tf.argmax(y_predict,2).eval() # )) #停止線程 coord.request_stop() coord.join(threads)