2.CNN图片多标签分类(基于TensorFlow实现验证码识别OCR)
原文:https://www.jianshu.com/p/596db72a7e00
0.1 2018.12.04 19:01* 字数 662 阅读 453评论 9喜欢 4
上一篇实现了图片CNN单标签分类(猫狗图片分类任务)
(地址:https://www.jianshu.com/p/47f0319028f2)
预告:下一篇用LSTM+CTC实现不定长文本的OCR,本质上是一种不固定标签个数的多标签分类问题
(地址:https://www.jianshu.com/p/7a2b227896a8)
本文所用到的10w验证码数据集百度网盘下载地址(也可使用下文代码自行生成):
https://pan.baidu.com/s/1N7bDHxIM38Vu7x9Z2Qr0og
利用本文代码训练并生成的模型(对应项目中的model文件夹):
https://pan.baidu.com/s/1GyEpLdM5FSxLYHSc6nq06w
项目简介:
(需要预先安装pip install captcha==0.1.1,pip install opencv-python,pip install flask, pip install tensorflow/pip install tensorflow-gpu)
本文采用CNN实现4位定长验证码图片OCR(生成的验证码固定由随机的4位大写字母组成),本质上是一张图片多个标签的分类问题(数据如下图所示)
1_CKAN.png
0_HZDZ.png
整体训练逻辑:
1,将图像传入到CNN中提取特征
2,将特征图拉伸输入到FC layer中得出分类预测向量
3,通过sigmoid交叉熵函数对预测向量和标签向量进行训练,得出最终模型(注意:多标签分类任务采用sigmoid,单标签分类采用softmax)
整体预测逻辑:
1,将图像传入到CNN(VGG16)中提取特征
2,将特征图拉伸输入到FC layer中得出分类预测向量
3,将预测向量做sigmoid操作,由于验证码固定是4位,所以将向量切分成4条,从每条中找到最大值,并映射到对应的字母上
制作成web服务:
利用flask框架将整个项目启动成web服务,使得项目支持http方式调用
启动服务后调用以下地址测试
http://127.0.0.1:5050/captchaOcr?img_path=./dataset/test/0_HZDZ.png
http://127.0.0.1:5050/captchaOcr?img_path=./dataset/test/1_CKAN.png
后续优化逻辑:
提取特征部分的CNN可以用RNN取代
本方案只能OCR固定长度文本,后续采用LSTM+CTC的方式来OCR非定长文本
运行命令:
自行生成验证码训练寄(本文生成了10w张,修改self.im_total_num变量):python CnnOcr.py create_dataset
对数据集进行训练:python CnnOcr.py train
对新的图片进行测试:python CnnOcr.py test
启动成http服务:python CnnOcr.py start
项目目录结构:
目录结构
训练过程:
epoch-1
epoch-2
epoch-3
epoch-4
epoch-5
epoch-6
epoch-6-结束
整体代码如下:
# coding:utf-8
from captcha.image import ImageCaptcha
import numpy as np
import cv2
import tensorflow as tf
import random, os, sys
from flask import request
from flask import Flask
import json
app = Flask(__name__)
class CnnOcr:
def __init__(self):
self.epoch_max = 6 # 最大迭代epoch次数
self.batch_size = 64 # 训练时每个批次参与训练的图像数目,显存不足的可以调小
self.lr = 1e-3 # 初始学习率
self.save_epoch = 1 # 每相隔多少个epoch保存一次模型
self.im_width = 128
self.im_height = 64
self.im_total_num = 100000 # 总共生成的验证码图片数量
self.train_max_num = self.im_total_num # 训练时读取的最大图片数目
self.val_num = 50 * self.batch_size # 不能大于self.train_max_num 做验证集用
self.words_num = 4 # 每张验证码图片上的数字个数
self.words = 'ABCDEFGHIJKLMNOPQRSTUVWXYZ'
self.label_num = self.words_num * len(self.words)
self.keep_drop = tf.placeholder(tf.float32)
self.x = None
self.y = None
def captchaOcr(self, img_path):
"""
验证码识别
:param img_path:
:return:
"""
im = cv2.imread(img_path)
im = cv2.resize(im, (self.im_width, self.im_height))
im = [im]
im = np.array(im, dtype=np.float32)
im -= 147
output = self.sess.run(self.max_idx_p, feed_dict={self.x: im, self.keep_drop: 1.})
ret = ''
for i in output.tolist()[0]:
ret = ret + self.words[int(i)]
return ret
def test(self, img_path):
"""
测试接口
:param img_path:
:return:
"""
self.x = tf.placeholder(tf.float32, [None, self.im_height, self.im_width, 3]) # 输入数据
self.pred = self.cnnNet()
self.output = tf.nn.sigmoid(self.pred)
self.predict = tf.reshape(self.pred, [-1, self.words_num, len(self.words)])
self.max_idx_p = tf.argmax(self.predict, 2)
saver = tf.train.Saver()
# tfconfig = tf.ConfigProto(allow_soft_placement=True)
# tfconfig.gpu_options.per_process_gpu_memory_fraction = 0.3 # 占用显存的比例
# self.ses = tf.Session(config=tfconfig)
self.sess = tf.Session()
self.sess.run(tf.global_variables_initializer()) # 全局tf变量初始化
# 加载w,b参数
saver.restore(self.sess, './model/CnnOcr-6')
im = cv2.imread(img_path)
im = cv2.resize(im, (self.im_width, self.im_height))
im = [im]
im = np.array(im, dtype=np.float32)
im -= 147
output = self.sess.run(self.max_idx_p, feed_dict={self.x: im, self.keep_drop: 1.})
ret = ''
for i in output.tolist()[0]:
ret = ret + self.words[int(i)]
print(ret)
def train(self):
x_train_list, y_train_list, x_val_list, y_val_list = self.getTrainDataset()
print('开始转换tensor队列')
x_train_list_tensor = tf.convert_to_tensor(x_train_list, dtype=tf.string)
y_train_list_tensor = tf.convert_to_tensor(y_train_list, dtype=tf.float32)
x_val_list_tensor = tf.convert_to_tensor(x_val_list, dtype=tf.string)
y_val_list_tensor = tf.convert_to_tensor(y_val_list, dtype=tf.float32)
x_train_queue = tf.train.slice_input_producer(tensor_list=[x_train_list_tensor], shuffle=False)
y_train_queue = tf.train.slice_input_producer(tensor_list=[y_train_list_tensor], shuffle=False)
x_val_queue = tf.train.slice_input_producer(tensor_list=[x_val_list_tensor], shuffle=False)
y_val_queue = tf.train.slice_input_producer(tensor_list=[y_val_list_tensor], shuffle=False)
train_im, train_label = self.dataset_opt(x_train_queue, y_train_queue)
train_batch = tf.train.batch(tensors=[train_im, train_label], batch_size=self.batch_size, num_threads=2)
val_im, val_label = self.dataset_opt(x_val_queue, y_val_queue)
val_batch = tf.train.batch(tensors=[val_im, val_label], batch_size=self.batch_size, num_threads=2)
print('开启训练')
self.learning_rate = tf.placeholder(dtype=tf.float32) # 动态学习率
self.x = tf.placeholder(tf.float32, [None, self.im_height, self.im_width, 3]) # 训练数据
self.y = tf.placeholder(tf.float32, [None, self.label_num]) # 标签
self.pred = self.cnnNet()
self.loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=self.pred, labels=self.y))
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate).minimize(self.loss)
self.predict = tf.reshape(self.pred, [-1, self.words_num, len(self.words)])
self.max_idx_p = tf.argmax(self.predict, 2)
self.y_predict = tf.reshape(self.y, [-1, self.words_num, len(self.words)])
self.max_idx_l = tf.argmax(self.y_predict, 2)
self.correct_pred = tf.equal(self.max_idx_p, self.max_idx_l)
self.accuracy = tf.reduce_mean(tf.cast(self.correct_pred, tf.float32))
with tf.Session() as self.sess:
# 全局tf变量初始化
self.sess.run(tf.global_variables_initializer())
coordinator = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=self.sess, coord=coordinator)
# 模型保存
saver = tf.train.Saver()
batch_max = len(x_train_list) // self.batch_size
total_step = 1
for epoch_num in range(self.epoch_max):
lr = self.lr * (1 - (epoch_num/self.epoch_max) ** 2) # 动态学习率
for batch_num in range(batch_max):
x_train_tmp, y_train_tmp = self.sess.run(train_batch)
# print(x_train_tmp.shape, y_train_tmp.shape)
# sys.exit()
self.sess.run(self.optimizer, feed_dict={self.x: x_train_tmp, self.y: y_train_tmp, self.learning_rate: lr, self.keep_drop: .5})
# 输出评价标准
if total_step % 50 == 0 or total_step == 1:
print()
print('epoch:%d/%d batch:%d/%d step:%d lr:%.10f' % ((epoch_num + 1), self.epoch_max, (batch_num + 1), batch_max, total_step, lr))
# 输出训练集评价
train_loss, train_acc = self.sess.run([self.loss, self.accuracy], feed_dict={self.x: x_train_tmp, self.y: y_train_tmp, self.keep_drop: 1.})
print('train_loss:%.10f train_acc:%.10f' % (np.mean(train_loss), train_acc))
# 输出验证集评价
val_loss_list, val_acc_list = [], []
for i in range(int(self.val_num/self.batch_size)):
x_val_tmp, y_val_tmp = self.sess.run(val_batch)
val_loss, val_acc = self.sess.run([self.loss, self.accuracy], feed_dict={self.x: x_val_tmp, self.y: y_val_tmp, self.keep_drop: 1.})
val_loss_list.append(np.mean(val_loss))
val_acc_list.append(np.mean(val_acc))
print(' val_loss:%.10f val_acc:%.10f' % (np.mean(val_loss), np.mean(val_acc)))
total_step += 1
# 保存模型
if (epoch_num + 1) % self.save_epoch == 0:
print('正在保存模型:')
saver.save(self.sess, './model/CnnOcr', global_step=(epoch_num + 1))
coordinator.request_stop()
coordinator.join(threads)
def cnnNet(self):
"""
cnn网络
:return:
"""
weight = {
# 输入 128*64*3
# 第一层
'wc1_1': tf.get_variable('wc1_1', [5, 5, 3, 32]), # 卷积 输出:128*64*32
'wc1_2': tf.get_variable('wc1_2', [5, 5, 32, 32]), # 卷积 输出:128*64*32
# 池化 输出:64*32*32
# 第二层
'wc2_1': tf.get_variable('wc2_1', [5, 5, 32, 64]), # 卷积 输出:64*32*64
'wc2_2': tf.get_variable('wc2_2', [5, 5, 64, 64]), # 卷积 输出:64*32*64
# 池化 输出:32*16*64
# 第三层
'wc3_1': tf.get_variable('wc3_1', [3, 3, 64, 64]), # 卷积 输出:32*16*256
'wc3_2': tf.get_variable('wc3_2', [3, 3, 64, 64]), # 卷积 输出:32*16*256
# 池化 输出:16*8*256
# 第四层
'wc4_1': tf.get_variable('wc4_1', [3, 3, 64, 64]), # 卷积 输出:16*8*64
'wc4_2': tf.get_variable('wc4_2', [3, 3, 64, 64]), # 卷积 输出:16*8*64
# 池化 输出:8*4*64
# 全链接第一层
'wfc_1': tf.get_variable('wfc_1', [8*4*64, 2048]),
# 全链接第二层
'wfc_2': tf.get_variable('wfc_2', [2048, 2048]),
# 全链接第三层
'wfc_3': tf.get_variable('wfc_3', [2048, self.label_num]),
}
biase = {
# 第一层
'bc1_1': tf.get_variable('bc1_1', [32]),
'bc1_2': tf.get_variable('bc1_2', [32]),
# 第二层
'bc2_1': tf.get_variable('bc2_1', [64]),
'bc2_2': tf.get_variable('bc2_2', [64]),
# 第三层
'bc3_1': tf.get_variable('bc3_1', [64]),
'bc3_2': tf.get_variable('bc3_2', [64]),
# 第四层
'bc4_1': tf.get_variable('bc4_1', [64]),
'bc4_2': tf.get_variable('bc4_2', [64]),
# 全链接第一层
'bfc_1': tf.get_variable('bfc_1', [2048]),
# 全链接第二层
'bfc_2': tf.get_variable('bfc_2', [2048]),
# 全链接第三层
'bfc_3': tf.get_variable('bfc_3', [self.label_num]),
}
# 第一层
net = tf.nn.conv2d(self.x, weight['wc1_1'], [1, 1, 1, 1], 'SAME') # 卷积
net = tf.nn.bias_add(net, biase['bc1_1'])
net = tf.nn.relu(net) # 加b 然后 激活
print('conv1', net)
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID') # 池化
print('pool1', net)
# 第二层
net = tf.nn.conv2d(net, weight['wc2_1'], [1, 1, 1, 1], padding='SAME') # 卷积
net = tf.nn.bias_add(net, biase['bc2_1'])
net = tf.nn.relu(net) # 加b 然后 激活
print('conv2', net)
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID') # 池化
print('pool2', net)
# 第三层
net = tf.nn.conv2d(net, weight['wc3_1'], [1, 1, 1, 1], padding='SAME') # 卷积
net = tf.nn.bias_add(net, biase['bc3_1'])
net = tf.nn.relu(net) # 加b 然后 激活
print('conv3', net)
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID') # 池化
print('pool3', net)
# 第四层
net = tf.nn.conv2d(net, weight['wc4_1'], [1, 1, 1, 1], padding='SAME') # 卷积
net = tf.nn.bias_add(net, biase['bc4_1'])
net = tf.nn.relu(net) # 加b 然后 激活
print('conv4', net)
net = tf.nn.max_pool(net, [1, 2, 2, 1], [1, 2, 2, 1], padding='VALID') # 池化
print('pool4', net)
# 拉伸flatten,把多个图片同时分别拉伸成一条向量
net = tf.reshape(net, shape=[-1, weight['wfc_1'].get_shape()[0]])
print('拉伸flatten', net)
# 全链接层
# fc第一层
net = tf.matmul(net, weight['wfc_1']) + biase['bfc_1']
net = tf.nn.dropout(net, self.keep_drop)
net = tf.nn.relu(net)
print('fc第一层', net)
# fc第二层
net = tf.matmul(net, weight['wfc_2']) + biase['bfc_2']
net = tf.nn.dropout(net, self.keep_drop)
net = tf.nn.relu(net)
print('fc第二层', net)
# fc第三层
net = tf.matmul(net, weight['wfc_3']) + biase['bfc_3']
print('fc第三层', net)
return net
def getTrainDataset(self):
"""
整理数据集,把图像resize为128*64*3,训练集做成self.im_total_num*128*64*3,把label做成0,1向量形式
:return:
"""
train_data_list = os.listdir('./dataset/train/')
print('共有%d张训练图片, 读取%d张:' % (len(train_data_list), self.train_max_num))
random.shuffle(train_data_list) # 打乱顺序
y_val_list, y_train_list = [], []
x_val_list = train_data_list[:self.val_num]
for x_val in x_val_list:
words_tmp = x_val.split('.')[0].split('_')[1]
y_val_list.append([1 if _w == w else 0 for w in words_tmp for _w in self.words])
x_train_list = train_data_list[self.val_num:self.train_max_num]
for x_train in x_train_list:
words_tmp = x_train.split('.')[0].split('_')[1]
y_train_list.append([1 if _w == w else 0 for w in words_tmp for _w in self.words])
return x_train_list, y_train_list, x_val_list, y_val_list
def createCaptchaDataset(self):
"""
生成训练用图片数据集
:return:
"""
image = ImageCaptcha(width=self.im_width, height=self.im_height, font_sizes=(56,))
for i in range(self.im_total_num):
words_tmp = ''
for j in range(self.words_num):
words_tmp = words_tmp + random.choice(self.words)
print(words_tmp, type(words_tmp))
im_path = './dataset/train/%d_%s.png' % (i, words_tmp)
print(im_path)
image.write(words_tmp, im_path)
return True
def dataset_opt(self, x_train_queue, y_train_queue):
"""
处理图片和标签
:param queue:
:return:
"""
queue = x_train_queue[0]
contents = tf.read_file('./dataset/train/' + queue)
im = tf.image.decode_jpeg(contents)
im = tf.image.resize_images(images=im, size=[self.im_height, self.im_width])
im = tf.reshape(im, tf.stack([self.im_height, self.im_width, 3]))
im -= 147 # 去均值化
# im /= 255 # 将像素处理在0~1之间,加速收敛
# im -= 0.5 # 将像素处理在-0.5~0.5之间
return im, y_train_queue[0]
if __name__ == '__main__':
opt_type = sys.argv[1:][0]
instance = CnnOcr()
if opt_type == 'create_dataset':
instance.createCaptchaDataset()
elif opt_type == 'train':
instance.train()
elif opt_type == 'test':
instance.test('./dataset/test/0_HZDZ.png')
elif opt_type == 'start':
# 将session持久化到内存中
instance.test('./dataset/test/0_HZDZ.png')
# 启动web服务
# http://127.0.0.1:5050/captchaOcr?img_path=./dataset/test/2_SYVD.png
@app.route('/captchaOcr', methods=['GET'])
def captchaOcr():
img_path = request.args.to_dict().get('img_path')
print(img_path)
ret = instance.captchaOcr(img_path)
print(ret)
return json.dumps({'img_path': img_path, 'ocr_ret': ret})
app.run(host='0.0.0.0', port=5050, debug=False)
4楼 · 2018.12.12 18:32
如果是sigmoid多标签分类的话,网络输出应该是26*1大小的向量,设置阈值,让向量里不止一个元素取1吧。现在26*4的话,就是把4个字符每一个字符做了onehot编码,用softmax输出26里最大概率的那个作为标签?不知道我这么理解对不对?
赞 回复
周小叨:
你的思路是对的,有个小问题,
我这边的FC层最后输出104*1的非one-hot向量,正好与y_label对应,在不做切分的情况下可以使用
sigmoid_cross_entropy_with_logits()函数直接计算交叉熵,softmax_cross_entropy_with_logits()无法计算非one-hot格式,如果切分计算就需要让4个loss最小,可以做到,但是有点麻烦了。另外如果只是单纯的26个标签的多标签分类,那么label就是26个元素的一维向量了,这样会导致区分不出ABCD和ACBD的区别;这个任务是4位的验证码识别,其实隐含了一个顺序问题,所以理解为104个固定顺序标签的多标签分类任务;过几天我会再发一篇LSTM+CTC的文章
2018.12.13 11:00 回复
huim:
@周小叨 我没考虑到顺序问题,现在理解了,蟹蟹~LSTM+CTC我做过,哈哈,有些小问题还没解决,等你更新了我再找你交流~
还有y的label是不是用的one-hot编码?小白我最近做多标签很多问题没想明白,求指教~
赞 回复
上一篇文章的单标签分类中,label是one-hot编码的向量([0,0,0,0,1]表示是第5类)
这篇多标签分类中,label不是one-hot格式,具体由以下代码获得:
[1 if _w == w else 0 for w in words_tmp for _w in self.words]
效果是:ABCD <<==>> [1,0,0,0......0, 0,1,0,0......0, 0,0,1,0......0, 0,0,0,1......0,]
向量长度是 26 * 4 = 104位
2楼 · 2018.12.09 21:38
你好,我想问一下为什么多标签分类要用sigmiod函数呢
赞 回复
周小叨:
假设最后一层输出10个值
可以用一个简单的概念来理解:
softmax处理之后10个值相加等于1,所以适用于单标签分类,比如第1类得分0.95,其余9类加起来只有0.05,那么分为第1类
sigmoid处理之后10个值每个值得分都在0~1之间,适用于多标签分类,比如第1个标签得分0.95,第4个标签得分0.99,其余8个标签得分均低于0.05,那么最终就输出1,4两个标签
2018.12.12 16:36 回复
月宸1.CNN图片单标签分类(基于TensorFlow实现基础VGG16网络)
本文所使用的开源数据集(kaggle猫狗大战):https://www.kaggle.com/c/dogs-vs-cats国内百度网盘下载地址:https://pan.baidu.com/s/12ab32UNYI_6o4i9HyX-n8Q 利用本文代码训练并生成的模型(对应...