迁移学习实现验证码预测

之前借助简单搭建的cnn模型执行验证码的分类任务训练，https://blog.csdn.net/weixin_41044499/article/details/94382397，由于缺乏训练多层神经网络的设备和条件，准确率并不高。这里通过借助vgg19的模型来抽取验证码图片的特征，再增加一个全连接层，做简单的分类问题。

相比多种调取现有cnn模型的方案，使用keras的vgg16直接操作的方式，安装和使用最为方便。

方法参考了这篇文章 https://blog.csdn.net/data_scientist/article/details/79041240

# -*- coding:utf-8 -*-
# /usr/bin/python
from keras.applications.vgg19 import VGG19
from keras.preprocessing import image
from keras.applications.vgg19 import preprocess_input
from keras.models import Model
import numpy as np

base_model = VGG19(weights='imagenet')
for layer in base_model.layers:
    print(layer.name)
model = Model(inputs=base_model.input, outputs=base_model.get_layer('predictions').output)
data = np.empty((300,1000),dtype="float32")
y=[]

##打开图片
for index,i in enumerate('ABC'):
    print(index)
    for j in range(0,100):
        img = image.load_img("yanzhengma2/"+i+'-'+str(j)+'.jpg', target_size=(224, 224))
        x = image.img_to_array(img)
        x = np.expand_dims(x, axis=0)
        x = preprocess_input(x)
        block4_pool_features = model.predict(x)
        ##把图片转换成数组形式
        data[index*100+j] = np.asarray(block4_pool_features,dtype="float32")
        y.append(index)

from sklearn import preprocessing
enc = preprocessing.OneHotEncoder()
enc.fit(np.array(y).reshape(-1, 1))    # fit来学习编码
y_new = enc.transform(np.array(y).reshape(-1, 1)).toarray()
np.savetxt("a.txt", data) # 缺省按照'%.18e'格式保存数据，以空格分隔
np.savetxt("b.txt", y_new,fmt="%d")

将图片像素经过vgg的转化后参数，保存到txt中。

# -*- coding:utf-8 -*-
# /usr/bin/python
import numpy as np
x = np.loadtxt("a.txt")
y = np.loadtxt("b.txt")
# np.savetxt("a.txt", data) # 缺省按照'%.18e'格式保存数据，以空格分隔
# np.savetxt("b.txt", y_new)
import tensorflow as tf
import numpy as np
## 原始数据
with tf.name_scope('data'):
    x = np.loadtxt("a.txt")
    y = np.loadtxt("b.txt")
## 创建参数，嵌套关系
with tf.name_scope('parameters'):
##权重
    with tf.name_scope('weights'):
        weight = tf.Variable(tf.random_uniform([1000,3],-1.0,1.0))
        tf.summary.histogram('weight',weight)
##偏置

    with tf.name_scope('biases'):
        bias = tf.Variable(tf.zeros([1,3]))
        tf.summary.histogram('bias',bias)
##get y_prediction
with tf.name_scope('y_prediction'):
    predict_y = tf.nn.softmax(tf.matmul(tf.cast(x,tf.float32), weight) + bias)

##compute the loss
with tf.name_scope('loss'):
    loss =tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=predict_y, labels=y))
    tf.summary.scalar('loss',loss)
##creat optimizer
optimizer = tf.train.AdamOptimizer(learning_rate=1e-3)
#creat train ,minimize the loss 
with tf.name_scope('train'):
    train = optimizer.minimize(loss)
#creat init
with tf.name_scope('init'):
    init = tf.global_variables_initializer()
##creat a Session 
sess = tf.Session()
#merged
merged = tf.summary.merge_all()
##initialize
writer = tf.summary.FileWriter("logs/", sess.graph)
sess.run(init)
## Loop
for step in range(1000):
    sess.run(train)
    rs=sess.run(merged)
    writer.add_summary(rs, step)
    if step % 25 == 0:
        correct_prediction = tf.equal(tf.argmax(predict_y, 1), tf.argmax(y, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        accuracy_value = sess.run(accuracy)
        print('step:%d accuracy:%.4f' % (step, accuracy_value))

对获得转化后的1000个变量，只加了一层全连接，经过多轮迭代后，准确率达到了100%，说明迁移学习的作用很大，用现有模型做特征抽取很有效果。