原文：www.e4rl.com/archives/198.html
原文已无法访问 ，所以特意从百度快照搬运过来。

所用项目及服务器

ThinkPHP：github.com/top-think/framework
think-captcha：github.com/top-think/think-captcha
captcha_trainer （训练）：github.com/kerlomz/captcha_trainer
captcha_platform （部署）：github.com/kerlomz/captcha_platfor...
MuggleOCR：pypi.org/project/muggle-ocr
captcha_trainer 作者介绍：www.jianshu.com/p/80ef04b16efc
矩池云 GPU服务器：matpool.com/

captcha_trainer 使用过程

1. 使用 ThinkPHP 调用 think-captcha 生成验证码

   composer create-project topthink/think tp
   cd tp
   composer require topthink/think-captcha
   php think run

2. 修改 think-captcha 源码，将验证码存储到 Session
   在 vendor/topthink/think-captcha/src/Captcha.php:195 插入以下代码

   Session::set('captcha', implode('', $code), '');

3. 在控制器内打印验证码

   public function code() {
    $code = Session::get('captcha', '');
    echo $code;
   }

4. 使用 Python 抓取图片验证码样本，按照 captcha_trainer 默认规则重命名保存

   import requests
   import threading
   import os
   import hashlib
   #
   def md5(s, salt=''):
    new_s = str(s) + salt
    m = hashlib.md5(new_s.encode())
    return m.hexdigest()
   #
   def get_captcha():
    session = requests.session()
    for i in range(0, 100000):
        try:
            content = session.get('http://x.com/captcha?'+str(i))
            if content.status_code != 200:
                continue
            code = session.get('http://x.com/index/Home/code')
            if code.status_code != 200:
                continue
            filename = '{}_{}.png'.format(code.text, md5(content.content))
            with open(os.path.join('captcha_images', filename), 'wb') as f:
                f.write(content.content)
                f.close()
        except Exception as e:
            print(str(e))
   #
   for i in range(1, 20):
    t = threading.Thread(target=get_captcha, args=())
    t.start()

5. 租用矩池云 RTX 2080 Ti 的GPU服务器进行训练
   

6. 使用 captcha_trainer 进行训练
   配置训练环境及项目

   pip3 install -r requirements.txt
   pip3 install tensorflow-gpu # tensorflow模块需独立安装
   mkdir -p projects/{project name}/
   vi projects/{project name}/model.yaml

model.yaml 模板参数介绍请参考 captcha_trainer 项目介绍，可以用 Windows 编译版生成后上传至服务器，需保持项目名一致。

项目名 TP-CNNX-GRU-H64-CTC-C1
./projects/TP-CNNX-GRU-H64-CTC-C1/model.yaml

# - requirement.txt  -  GPU: tensorflow-gpu, CPU: tensorflow
# - If you use the GPU version, you need to install some additional applications.
System:
  MemoryUsage: 0.8
  Version: 2

# CNNNetwork: [CNN5, ResNet, DenseNet]
# RecurrentNetwork: [CuDNNBiLSTM, CuDNNLSTM, CuDNNGRU, BiLSTM, LSTM, GRU, BiGRU, NoRecurrent]
# - The recommended configuration is CNN5+GRU
# UnitsNum: [16, 64, 128, 256, 512]
# - This parameter indicates the number of nodes used to remember and store past states.
# Optimizer: Loss function algorithm for calculating gradient.
# - [AdaBound, Adam, Momentum]
# OutputLayer: [LossFunction, Decoder]
# - LossFunction: [CTC, CrossEntropy]
# - Decoder: [CTC, CrossEntropy]
NeuralNet:
  CNNNetwork: CNNX
  RecurrentNetwork: GRU
  UnitsNum: 64
  Optimizer: RAdam
  OutputLayer:
    LossFunction: CTC
    Decoder: CTC

# ModelName: Corresponding to the model file in the model directory
# ModelField: [Image, Text]
# ModelScene: [Classification]
# - Currently only Image-Classification is supported.
Model:
  ModelName: TP-CNNX-GRU-H64-CTC-C1
  ModelField: Image
  ModelScene: Classification

# FieldParam contains the Image, Text.
# When you filed to Image:
# - Category: Provides a default optional built-in solution:
# -- [ALPHANUMERIC, ALPHANUMERIC_LOWER, ALPHANUMERIC_UPPER,
# -- NUMERIC, ALPHABET_LOWER, ALPHABET_UPPER, ALPHABET, ALPHANUMERIC_CHS_3500_LOWER]
# - or can be customized by:
# -- ['Cat', 'Lion', 'Tiger', 'Fish', 'BigCat']
# - Resize: [ImageWidth, ImageHeight/-1, ImageChannel]
# - ImageChannel: [1, 3]
# - In order to automatically select models using image size, when multiple models are deployed at the same time:
# -- ImageWidth: The width of the image.
# -- ImageHeight: The height of the image.
# - MaxLabelNum: You can fill in -1, or any integer, where -1 means not defining the value.
# -- Used when the number of label is fixed
# When you filed to Text:
# This type is temporarily not supported.
FieldParam:
  Category: ['2', '3', '4', '5', '6', '7', '8', 'a', 'b', 'c', 'd', 'e', 'f', 'h', 'i', 'j', 'k', 'm', 'n', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'J', 'K', 'L', 'M', 'N', 'P', 'Q', 'R', 'T', 'U', 'V', 'W', 'X', 'Y']
  Resize: [224, 70]
  ImageChannel: 1
  ImageWidth: 224
  ImageHeight: 70
  MaxLabelNum: 4
  OutputSplit: 
  AutoPadding: True

# The configuration is applied to the label of the data source.
# LabelFrom: [FileName, XML, LMDB]
# ExtractRegex: Only for methods extracted from FileName:
# - Default matching apple_20181010121212.jpg file.
# - The Default is .*?(?=_.*\.)
# LabelSplit: Only for methods extracted from FileName:
# - The split symbol in the file name is like: cat&big cat&lion_20181010121212.png
# - The Default is null.
Label:
  LabelFrom: FileName
  ExtractRegex: .*?(?=_)
  LabelSplit: 

# DatasetPath: [Training/Validation], The local absolute path of a packed training or validation set.
# SourcePath:  [Training/Validation], The local absolute path to the source folder of the training or validation set.
# ValidationSetNum: This is an optional parameter that is used when you want to extract some of the validation set
# - from the training set when you are not preparing the validation set separately.
# SavedSteps: A Session.run() execution is called a Step,
# - Used to save training progress, Default value is 100.
# ValidationSteps: Used to calculate accuracy, Default value is 500.
# EndAcc: Finish the training when the accuracy reaches [EndAcc*100]% and other conditions.
# EndCost: Finish the training when the cost reaches EndCost and other conditions.
# EndEpochs: Finish the training when the epoch is greater than the defined epoch and other conditions.
# BatchSize: Number of samples selected for one training step.
# ValidationBatchSize: Number of samples selected for one validation step.
# LearningRate: [0.1, 0.01, 0.001, 0.0001]
# - Use a smaller learning rate for fine-tuning.
Trains:
  DatasetPath:
    Training: 
      - ./projects/TP-CNNX-GRU-H64-CTC-C1/dataset/Trains.0.tfrecords
    Validation: 
      - ./projects/TP-CNNX-GRU-H64-CTC-C1/dataset/Validation.0.tfrecords
  SourcePath:
    Training: 
      - /root/captcha_images
    Validation: 
  ValidationSetNum: 300
  SavedSteps: 100
  ValidationSteps: 500
  EndAcc: 0.95
  EndCost: 0.5
  EndEpochs: 2
  BatchSize: 64
  ValidationBatchSize: 300
  LearningRate: 0.001

# Binaryzation: The argument is of type list and contains the range of int values, -1 is not enabled.
# MedianBlur: The parameter is an int value, -1 is not enabled.
# GaussianBlur: The parameter is an int value, -1 is not enabled.
# EqualizeHist: The parameter is an bool value.
# Laplace: The parameter is an bool value.
# WarpPerspective: The parameter is an bool value.
# Rotate: The parameter is a positive integer int type greater than 0, -1 is not enabled.
# PepperNoise: This parameter is a float type less than 1, -1 is not enabled.
# Brightness: The parameter is an bool value.
# Saturation: The parameter is an bool value.
# Hue: The parameter is an bool value.
# Gamma: The parameter is an bool value.
# ChannelSwap: The parameter is an bool value.
# RandomBlank: The parameter is a positive integer int type greater than 0, -1 is not enabled.
# RandomTransition: The parameter is a positive integer int type greater than 0, -1 is not enabled.
DataAugmentation:
  Binaryzation: -1
  MedianBlur: -1
  GaussianBlur: -1
  EqualizeHist: False
  Laplace: False
  WarpPerspective: False
  Rotate: -1
  PepperNoise: -1.0
  Brightness: False
  Saturation: False
  Hue: False
  Gamma: False
  ChannelSwap: False
  RandomBlank: -1
  RandomTransition: -1
  RandomCaptcha: 
     Enable: False
     FontPath: 

# Binaryzation: The parameter is an integer number between 0 and 255, -1 is not enabled.
# ReplaceTransparent: Transparent background replacement, bool type.
# HorizontalStitching: Horizontal stitching, bool type.
# ConcatFrames: Horizontally merge two frames according to the provided frame index list, -1 is not enabled.
# BlendFrames: Fusion corresponding frames according to the provided frame index list, -1 is not enabled.
# - [-1] means all frames
Pretreatment:
  Binaryzation: -1
  ReplaceTransparent: True
  HorizontalStitching: False
  ConcatFrames: -1
  BlendFrames: -1
  ExecuteMap: {}

打包训练集后开始训练

python3 make_dataset.py TP-CNNX-GRU-H64-CTC-C1
python3 trains.py TP-CNNX-GRU-H64-CTC-C1

7. Linux 训练环境强制结束任务
   因为 think-captcha 随机调用系统字体生成图片验证码，部分字体显示只有大写字母，与 think-captcha 生成的验证码存在大小写偏差，导致 captcha_trainer 训练正确率只能维持在 0.6 左右，无法满足结束训练任务的 0.95 正确率，但忽略大小写后的正确率已满足需求，所以需要强制结束训练任务，编译模型。

Windows 环境下可以直接点击 Stop 按钮结束任务后再点击 Compile 按钮编译模型。

Linux 环境结束任务需要修改 trains.py 代码，简单分析后发现 trains.py 共有两处正确率判断

# trains.py:308
# 满足终止条件但尚未完成当前epoch时跳出epoch循环
if self.achieve_cond(acc=accuracy, cost=batch_cost, epoch=epoch_count):
    break

# trains.py:314
if self.achieve_cond(acc=accuracy, cost=batch_cost, epoch=epoch_count):
    # sess.close()
    tf.compat.v1.keras.backend.clear_session()
    sess.close()
    self.compile_graph(accuracy)
    tf.compat.v1.logging.info('Total Time: {} sec.'.format(time.time() - start_time))

找到代码后直接简单粗暴的把判断条件改成 1==1 让条件成立即可结束任务

# trains.py:308
# 满足终止条件但尚未完成当前epoch时跳出epoch循环
if 1==1:
    break

# trains.py:314
if 1==1:
    # sess.close()
    tf.compat.v1.keras.backend.clear_session()
    sess.close()
    self.compile_graph(accuracy)
    tf.compat.v1.logging.info('Total Time: {} sec.'.format(time.time() - start_time))

captcha_trainer 支持中断任务恢复，修改代码后按 Ctrl+C 结束任务，重新执行 python3 trains.py，初始化后将直接结束训练任务，编译模型。

部署验证码识别接口

使用 captcha_platform 项目进行 docker 部署

1. 将 captcha_trainer 训练后编译生成的模型复制到 captcha_platform 项目中

   mv -rf captcha_trainer/out/* captcha_platform/

2. 构建并启动 docker

   cd captcha_platform/
   docker build . 
   docker run -d -p 19952:19952 [image:tag]

识别测试

13万 think-captcha 图片验证码样本经过约 11 个小时的训练后，使用未经过训练的新样本进行识别测试，识别成功率在 95% 左右

使用 Python 脚本进行识别测试

import requests
import base64
import os
import re
#
dir = 'test-1' # 未经过训练的图片验证码样本目录
success = 0
error = 0
count = 0
file_list = os.listdir(dir)
#
for i in range(0, 1000):
    filename = file_list[i]
    origin = re.match(r'.*?(?=_.*\.)', filename).group()
    if not origin:
        continue
    src = os.path.join(dir, filename)
    img = base64.b64encode(open(src, 'rb').read())
    post_data = {
        'image': img,
        'model_name': 'TP-CNNX-GRU-H64-CTC-C1'
    }
    res = requests.post('http://localhost:19952/captcha/v1', data=post_data)
    json_res = res.json()
    # 忽略大小写
    if json_res['code'] == 0 and str(json_res['message']).upper() == origin.upper():
        print('Sample: {}, Identify: {}, Status: {}'.format(org, json_res['message'], 'success'))
        success += 1
    else:
        print('Sample: {}, Identify: {}, Status: {}'.format(org, json_res['message'], 'error'))
        error += 1
    count += 1
print('Count: {}, Success: {}, Error: {}'.format(count, success, error))

测试结果

模型下载

链接: pan.baidu.com/s/1e0quRSqMV8lP6XXoS... 提取码:sg6c

本作品采用《CC 协议》，转载必须注明作者和本文链接