使用深度学习模型进行图片验证码识别
基于深度学习模型进行图片验证码识别（keras）目的：在登录网站，尤其使用爬虫工具时通常会遇到验证码，大多数时候都是使用第三方付费的接口识别验证码结果，学习了深度学习后可以尝试自己训练一个验证码识别的模型。
过程：
# 1.构造数据集（这里通过python工具自行生成图片验证码）
### （1） 生成验证码训练、测试图片
import string
letters = string.digits+string.ascii_letters
list_labels =
def create_image_labels(path,num_img=100):
width = 160
height = 60
for i in range(num_img):
chars = ""
for __ in range(4):
chars += random.choice(list_labels)
image = ImageCaptcha().generate_image(chars)
img = image.convert('L') #转化为灰度图
resized_image = img.resize((width, height), Image.ANTIALIAS) #压缩图片
files = path+str(i)+"_" +chars+".png"
resized_image.save(files)
print("bulid success")
#create_image_labels("./image/",num_img=40000)
#create_image_labels("./test/",num_img=100)
### (2)numpy整理图片数据
def get_train_Dataset(file_name):
weight = 160
height = 60
files =os.listdir(file_name)
fileList =
labels = []
x = np.zeros()
dict_labels = {}
for index,label in enumerate(list_labels):
dict_labels = index
threshold = 195#阈值
table = []
for i in range(256):
if i < threshold:
table.append(0)
else:
table.append(1)
for index,f_img in enumerate(fileList):
#处理label
label = f_img.split("_")[-1].split(".")
label_split = for la in label]
labels.append(label_split)
#处理图片转换为为数组
img = cv2.imread(f_img,cv2.IMREAD_GRAYSCALE)
median = cv2.medianBlur(img, 5)
image = Image.fromarray(cv2.cvtColor(median,cv2.COLOR_BGR2RGB))
img= image.convert("L")#灰度
img = img.point(table, '1')#二值化
x_train = np.array(img).reshape(-1,height*weight)
# x_shape = x_train.shape
# x_train = x_train.reshape(x_shape*x_shape,x_shape).max(axis=1)
# trian = x_train/255
x = x_train
### (3)可保存整理后的数据npz
train_x,train_y = get_train_Dataset("./image")
test_x,test_y = get_train_Dataset("./test")
#保存为npz 文件(占用存储空间)
np.savez('./image_train_and_test06.npz',train_x=train_x,train_y=train_y,test_x=test_x,test_y=test_y)
# 2.构造模型训练及保存模型参数
### (1)模块导入
from keras.models import load_model
import numpy as np
from PIL import Image
from keras.models import Model
from keras.layers import Input,Dense,Convolution2D,MaxPooling2D,Flatten,Dropout
from keras.optimizers import Adam
from keras.utils import np_utils
from keras.utils import plot_model
#数据准备
CLASS_NUM = 62
weight = 160
height = 60
data = np.load("./image_train_and_test06.npz")
X_train, X_test,y_train, y_test =data["train_x"],data["test_x"],data["train_y"],data["test_y"]
### （2）构建模型
#搭建网络模型，
def main(X_train,y_train,epochs=50):
# 输入层
inputs = Input(shape=(height,weight,1))
x = inputs
x=Convolution2D(32,(3,3),activation='relu')(x)
x=Convolution2D(32,(3,3),activation='relu')(x)
x=MaxPooling2D((2,2))(x)
x=Dropout(0.5)(x)
x=Convolution2D(64,(3,3),activation='relu')(x)
x=Convolution2D(64,(3,3),activation='relu')(x)
x=MaxPooling2D((2,2))(x)
x=Dropout(0.5)(x)
x=Convolution2D(128,(3,3),activation='relu')(x)
x=MaxPooling2D((2,2))(x)
x=Flatten()(x)
x=Dropout(0.5)(x)
#卷积层及全连接层等相关层
x = Dense(1024, activation="relu")(x)
#多输出，定义两个输出，指定名字标识
fc_a=Dense(name='fc_a',units=CLASS_NUM,activation="softmax")(x)
fc_b=Dense(name='fc_b',units=CLASS_NUM,activation="softmax")(x)
fc_c=Dense(name='fc_c',units=CLASS_NUM,activation="softmax")(x)
fc_d=Dense(name='fc_d',units=CLASS_NUM,activation="softmax")(x)
#单输入多输出
model = Model(inputs=inputs, outputs=)
#目标函数定义，需与输出层名字对应
losses = {'fc_a': 'categorical_crossentropy',
'fc_b': 'categorical_crossentropy',
'fc_c': 'categorical_crossentropy',
'fc_d': 'categorical_crossentropy'}
model.compile(optimizer=Adam(0.00001,decay=1e-6,clipnorm=1.),
loss=losses,
metrics=['accuracy'])
X_train = X_train.reshape(-1,height,weight,1)/255
y_train1 = np_utils.to_categorical(y_train[:,0],num_classes=CLASS_NUM)
y_train2 = np_utils.to_categorical(y_train[:,1],num_classes=CLASS_NUM)
y_train3 = np_utils.to_categorical(y_train[:,2],num_classes=CLASS_NUM)
y_train4 = np_utils.to_categorical(y_train[:,3],num_classes=CLASS_NUM)
model.fit(X_train,,epochs=epochs,batch_size=64,shuffle=True)
return model
### （3）保存，加载模型
model = main(X_train,y_train,epochs=800)
#画结构图
#import os
#os.environ["PATH"] += os.pathsep + 'D:/software/graph/bin/'
#plot_model(model, to_file='./model4.png', show_shapes=True)
#保存模型
model.save("./my_morestring62_160_model04.h5")
#加载模型
#model = load_model("./my_morestring62_160_model04.h5")
# 3.预测
import cv2
def prep_image(path):
#加载模型
model = load_model("./my_morestring62_160_model04.h5")
#img = Image.open(path)
#img = img.resize((weight,height))
threshold = 215#阈值
table = []
for i in range(256):
if i < threshold:
table.append(0)
else:
table.append(1)
img = cv2.imread(path,cv2.IMREAD_GRAYSCALE)
median = cv2.medianBlur(result, 5)
image = Image.fromarray(cv2.cvtColor(median,cv2.COLOR_BGR2RGB))
img= image.convert("L")#灰度
img = img.point(table, '1')#二值化
img = img.resize((weight,height))
X_test = np.array(img).reshape(-1,height*weight)
X_test = X_test.reshape(-1,height,weight,1)/255
prep1,prep2,prep3,prep4 = model.predict(X_test)
prep = np.hstack((np.argmax(prep1,axis=1).reshape(-1,1),np.argmax(prep2,axis=1).reshape(-1,1),\
np.argmax(prep3,axis=1).reshape(-1,1),np.argmax(prep4,axis=1).reshape(-1,1))).reshape(-1,4)
import string
letters = string.digits+string.ascii_letters
list_labels =
num_dicts = {}
for index,label in enumerate(list_labels):
num_dicts = label
new_labels = []
for i in range(prep.shape):
labels = []
for j in range(prep.shape):
labels.append(num_dicts])
new_labels.append("".join(labels))
return new_labels
#result = prep_image("./vaild/0_t76j.png")
print(result)
上面模型结构为vgg 网络结构（训练结果不是很理想）
应该为先这个模型结构图
m58758788 发表于 2021-10-3 00:55
厉害了 可惜现在好多都是滑块 点选
现在很多第三方平台也支持点选的（使用 “图鉴“ 点选*06车票都不错；超级鹰也可以，不过要差一点），滑块可以模拟操作，或者逆向解析js，不过难度较大。 厉害了 可惜现在好多都是滑块 点选 第三方不知道行不行 滑块点选的识别问题不大，难的是如何通过人机检测吧 chinasmu 发表于 2021-10-3 08:25
滑块点选的识别问题不大，难的是如何通过人机检测吧
对的，比如某宝的滑块很容易模拟，但是账号识别为不同等级就比较狠了 牛逼{:301_993:}，都用到深度学习了，大佬{:301_995:} 厉害,感谢分享啊 识别率怎么样?
页:
[1]
2