Tensorflow搭建CNN实现验证码识别

​整个项目代码分为三部分:

  • Generrate_Captcha:

生成验证码图片(训练集,验证集和测试集);

读取图片数据和标签(标签即为图片文件名);

  • cnn_model:卷积神经网络;

  • driver:模型训练及评估。

1、配置项

class Config(object):
    width = 160  # 验证码图片的宽
    height = 60  # 验证码图片的高
    char_num = 4  # 验证码字符个数
    characters = range(10)  # 数字[0,9]
    test_folder = 'test'    # 测试集文件夹,下同
    train_folder = 'train'
    validation_folder = 'validation'
    tensorboard_folder = 'tensorboard'  # tensorboard的log路径
    generate_num = (5000, 500, 500)  # 训练集,验证集和测试集数量
    alpha = 1e-3  # 学习率
    Epoch = 100  # 训练轮次
    batch_size = 64     # 批次数量
    keep_prob = 0.5     # dropout比例
    print_per_batch = 20    # 每多少次输出结果
    save_per_batch = 20     # 每多少次写入tensorboard

2、生成验证码(class Generate)

验证码图片示例:

file

0478

check_path():检查文件夹是否存在,如不存在则创建。

gen_captcha():生成验证码方法,写入之前检查是否以存在,如存在重新生成。

3、读取数据(classs ReadData)

read_data():返回图片数组(numpy.array格式)和标签(即文件名);
label2vec():将文件名转为向量;


label = '1327'
​
label_vec = [0,1,0,0,0,0,0,0,0,0,
            0,0,0,1,0,0,0,0,0,0,
            0,0,1,0,0,0,0,0,0,0,
            0,0,0,0,0,0,0,1,0,0]

load_data():加载文件夹下所有图片,返回图片数组,标签和图片数量。

4、定义模型(cnn_model)

采用三层卷积,filter_size均为5,为避免过拟合,每层卷积后面均接dropout操作,最终将的图像转为的矩阵。
大致结构如下:

Tensorflow搭建CNN实现验证码识别_第1张图片

模型结构

5、训练&评估

  • next_batch():迭代器,分批次返还数据;

  • feed_data():给模型“喂”数据;

x:图像数组;
y:图像标签;
keep_prob:dropout比例;

  • evaluate():模型评估,用于验证集和测试集。

  • run_model():训练&评估

6、目前效果

目前经过4000次迭代训练集准确率可达99%以上,测试集准确率93%,还是存在一点过拟合,不过现在模型是基于CPU训练的,完成一次训练耗费时间大约4个小时左右,后续调整了再进行更新。

Images for train :10000, for validation : 1000, for test : 1000
Epoch : 1
Step     0, train_acc:   7.42%, train_loss:  1.43, val_acc:   9.85%, val_loss:  1.40, improved:*  
Step    20, train_acc:  12.50%, train_loss:  0.46, val_acc:  10.35%, val_loss:  0.46, improved:*  
Step    40, train_acc:   9.38%, train_loss:  0.37, val_acc:  10.10%, val_loss:  0.37, improved:   
Step    60, train_acc:   7.42%, train_loss:  0.34, val_acc:  10.25%, val_loss:  0.34, improved:   
Step    80, train_acc:   7.81%, train_loss:  0.33, val_acc:   9.82%, val_loss:  0.33, improved:   
Step   100, train_acc:  12.11%, train_loss:  0.33, val_acc:  10.00%, val_loss:  0.33, improved:   
Step   120, train_acc:   9.77%, train_loss:  0.33, val_acc:  10.07%, val_loss:  0.33, improved:   
Step   140, train_acc:   8.98%, train_loss:  0.33, val_acc:  10.40%, val_loss:  0.33, improved:*  
Epoch : 2
Step   160, train_acc:   8.20%, train_loss:  0.33, val_acc:  10.52%, val_loss:  0.33, improved:*  
...
Epoch : 51
Step  7860, train_acc: 100.00%, train_loss:  0.01, val_acc:  92.37%, val_loss:  0.08, improved:   
Step  7880, train_acc:  99.61%, train_loss:  0.01, val_acc:  92.28%, val_loss:  0.08, improved:   
Step  7900, train_acc: 100.00%, train_loss:  0.01, val_acc:  92.42%, val_loss:  0.08, improved:   
Step  7920, train_acc: 100.00%, train_loss:  0.00, val_acc:  92.83%, val_loss:  0.08, improved:   
Step  7940, train_acc: 100.00%, train_loss:  0.01, val_acc:  92.77%, val_loss:  0.08, improved:   
Step  7960, train_acc: 100.00%, train_loss:  0.01, val_acc:  92.68%, val_loss:  0.08, improved:   
Step  7980, train_acc: 100.00%, train_loss:  0.00, val_acc:  92.63%, val_loss:  0.09, improved:   
No improvement for over 1000 steps, auto-stopping....
Test accuracy:  93.00%, loss:  0.08

7、Tensorboard

每次训练之前将Tensorboard路径下的文件删除,不然趋势图上会凌乱。
Accurracy

Tensorflow搭建CNN实现验证码识别_第2张图片

loss

Tensorflow搭建CNN实现验证码识别_第3张图片
原文链接及完整代码:

https://github.com/AwesomeTang/Captcha_CNN

​文源网络,仅供学习之用,如有侵权,联系删除。

我将优质的技术文章和经验总结都汇集在了我的公众号【Python圈子】里,为方便大家学习,还整理了一套学习资料,免费提供给热爱Python的同学! 更有学习交流群,多交流问题才能更快进步~
Tensorflow搭建CNN实现验证码识别_第4张图片

你可能感兴趣的:(Tensorflow搭建CNN实现验证码识别)