dyboy2017 / captcha_check

CNN卷积神经网络训练识别验证码模型的一次尝试！

captcha_check

author: DYBOY

项目简介：

利用 captcha 库生成的图形验证码，并对其进行内容识别。首先通过对验证码图片进行一个“灰度”处理，使之变为灰度图，灰度图有利于去除杂色，便于提高模型精度，提升训练速度。训练过程中，使用了三层CNN，最终的准确率最高可达到98.75%

运行环境：

• python3.6
• tensorflow 1.10 (运行时，缺少什么就安装对应包 例如安装tensorflow：pip3 install tensorflow -i https://pypi.tuna.tsinghua.edu.cn/simple )
• windows 10 家庭版

USAGE:

• 训练模型：python3 model_train.py
• 测试模型：python3 model_test.py

效果：

项目介绍文章地址：

• https://blog.dyboy.cn/program/100.html

训练过程记录：

[root@VM_96_17_centos captcha]# /usr/local/bin/python3.6 model_train.py 
2018-11-26 09:29:57.038117: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-11-26 09:29:57.408852: W tensorflow/core/framework/allocator.cc:122] Allocation of 41943040 exceeds 10% of system memory.
2018-11-26 09:29:57.451278: W tensorflow/core/framework/allocator.cc:122] Allocation of 41943040 exceeds 10% of system memory.
2018-11-26 09:29:57.498523: W tensorflow/core/framework/allocator.cc:122] Allocation of 41943040 exceeds 10% of system memory.
2018-11-26 09:29:57.525206: W tensorflow/core/framework/allocator.cc:122] Allocation of 41943040 exceeds 10% of system memory.
2018-11-26 09:29:57.566340: W tensorflow/core/framework/allocator.cc:122] Allocation of 41943040 exceeds 10% of system memory.
Mon Nov 26 09:30:03 2018  step: 0  accuracy: 0.1
Mon Nov 26 09:33:47 2018  step: 100  accuracy: 0.1175
Mon Nov 26 09:37:29 2018  step: 200  accuracy: 0.1
Mon Nov 26 09:41:11 2018  step: 300  accuracy: 0.095
Mon Nov 26 09:44:53 2018  step: 400  accuracy: 0.1
Mon Nov 26 09:48:32 2018  step: 500  accuracy: 0.105
Mon Nov 26 09:52:11 2018  step: 600  accuracy: 0.0575
Mon Nov 26 09:55:50 2018  step: 700  accuracy: 0.09
Mon Nov 26 09:59:30 2018  step: 800  accuracy: 0.105
Mon Nov 26 10:03:10 2018  step: 900  accuracy: 0.085
Mon Nov 26 10:06:50 2018  step: 1000  accuracy: 0.0875
Mon Nov 26 10:10:31 2018  step: 1100  accuracy: 0.09
Mon Nov 26 10:14:11 2018  step: 1200  accuracy: 0.08
Mon Nov 26 10:17:51 2018  step: 1300  accuracy: 0.0825
Mon Nov 26 10:21:31 2018  step: 1400  accuracy: 0.135
Mon Nov 26 10:25:11 2018  step: 1500  accuracy: 0.18
Mon Nov 26 10:28:51 2018  step: 1600  accuracy: 0.31
Mon Nov 26 10:32:31 2018  step: 1700  accuracy: 0.4225
Mon Nov 26 10:36:09 2018  step: 1800  accuracy: 0.4625
Mon Nov 26 10:39:49 2018  step: 1900  accuracy: 0.4925
Mon Nov 26 10:43:28 2018  step: 2000  accuracy: 0.5475
Mon Nov 26 10:47:07 2018  step: 2100  accuracy: 0.58
Mon Nov 26 10:50:46 2018  step: 2200  accuracy: 0.6725
Mon Nov 26 10:54:25 2018  step: 2300  accuracy: 0.69
Mon Nov 26 10:58:06 2018  step: 2400  accuracy: 0.735
Mon Nov 26 11:01:46 2018  step: 2500  accuracy: 0.735
Mon Nov 26 11:05:27 2018  step: 2600  accuracy: 0.745
Mon Nov 26 11:09:07 2018  step: 2700  accuracy: 0.825
Mon Nov 26 11:12:46 2018  step: 2800  accuracy: 0.7675
Mon Nov 26 11:16:26 2018  step: 2900  accuracy: 0.8425
Mon Nov 26 11:20:05 2018  step: 3000  accuracy: 0.825
Mon Nov 26 11:23:43 2018  step: 3100  accuracy: 0.855
Mon Nov 26 11:27:23 2018  step: 3200  accuracy: 0.8425
Mon Nov 26 11:31:02 2018  step: 3300  accuracy: 0.8375
Mon Nov 26 11:34:41 2018  step: 3400  accuracy: 0.8575
Mon Nov 26 11:38:20 2018  step: 3500  accuracy: 0.8475
Mon Nov 26 11:41:59 2018  step: 3600  accuracy: 0.8775
Mon Nov 26 11:45:38 2018  step: 3700  accuracy: 0.8825
Mon Nov 26 11:49:17 2018  step: 3800  accuracy: 0.9025
Mon Nov 26 11:52:57 2018  step: 3900  accuracy: 0.8625
Mon Nov 26 11:56:37 2018  step: 4000  accuracy: 0.92
Mon Nov 26 12:00:17 2018  step: 4100  accuracy: 0.8925
Mon Nov 26 12:03:56 2018  step: 4200  accuracy: 0.9075
Mon Nov 26 12:07:35 2018  step: 4300  accuracy: 0.91
Mon Nov 26 12:11:15 2018  step: 4400  accuracy: 0.92
Mon Nov 26 12:14:54 2018  step: 4500  accuracy: 0.915
Mon Nov 26 12:18:36 2018  step: 4600  accuracy: 0.9275
Mon Nov 26 12:22:16 2018  step: 4700  accuracy: 0.9375
Mon Nov 26 12:25:56 2018  step: 4800  accuracy: 0.905
Mon Nov 26 12:29:36 2018  step: 4900  accuracy: 0.915
Mon Nov 26 12:33:17 2018  step: 5000  accuracy: 0.945
Mon Nov 26 12:36:56 2018  step: 5100  accuracy: 0.9475
Mon Nov 26 12:40:35 2018  step: 5200  accuracy: 0.9375
Mon Nov 26 12:44:15 2018  step: 5300  accuracy: 0.9675
Mon Nov 26 12:47:57 2018  step: 5400  accuracy: 0.945
Mon Nov 26 12:51:37 2018  step: 5500  accuracy: 0.96
Mon Nov 26 12:55:16 2018  step: 5600  accuracy: 0.93
Mon Nov 26 12:58:56 2018  step: 5700  accuracy: 0.9
Mon Nov 26 13:02:36 2018  step: 5800  accuracy: 0.9425
Mon Nov 26 13:06:15 2018  step: 5900  accuracy: 0.9275
Mon Nov 26 13:09:54 2018  step: 6000  accuracy: 0.9225
Mon Nov 26 13:13:34 2018  step: 6100  accuracy: 0.945
Mon Nov 26 13:17:13 2018  step: 6200  accuracy: 0.945
Mon Nov 26 13:20:52 2018  step: 6300  accuracy: 0.9475
Mon Nov 26 13:24:32 2018  step: 6400  accuracy: 0.9675
Mon Nov 26 13:28:13 2018  step: 6500  accuracy: 0.95
Mon Nov 26 13:31:51 2018  step: 6600  accuracy: 0.9375
Mon Nov 26 13:35:30 2018  step: 6700  accuracy: 0.9475
Mon Nov 26 13:39:09 2018  step: 6800  accuracy: 0.95
Mon Nov 26 13:42:49 2018  step: 6900  accuracy: 0.96
Mon Nov 26 13:46:29 2018  step: 7000  accuracy: 0.955
Mon Nov 26 13:50:09 2018  step: 7100  accuracy: 0.9425
Mon Nov 26 13:53:48 2018  step: 7200  accuracy: 0.9625
Mon Nov 26 13:57:27 2018  step: 7300  accuracy: 0.94
Mon Nov 26 14:01:10 2018  step: 7400  accuracy: 0.97
Mon Nov 26 14:04:52 2018  step: 7500  accuracy: 0.9675
Mon Nov 26 14:08:31 2018  step: 7600  accuracy: 0.9575
Mon Nov 26 14:12:11 2018  step: 7700  accuracy: 0.96
Mon Nov 26 14:15:51 2018  step: 7800  accuracy: 0.9425
Mon Nov 26 14:19:31 2018  step: 7900  accuracy: 0.965
Mon Nov 26 14:23:11 2018  step: 8000  accuracy: 0.9725
Mon Nov 26 14:26:51 2018  step: 8100  accuracy: 0.9825
Mon Nov 26 14:30:32 2018  step: 8200  accuracy: 0.975
Mon Nov 26 14:34:11 2018  step: 8300  accuracy: 0.9625
Mon Nov 26 14:37:49 2018  step: 8400  accuracy: 0.9775
Mon Nov 26 14:41:29 2018  step: 8500  accuracy: 0.955
Mon Nov 26 14:45:09 2018  step: 8600  accuracy: 0.9675
Mon Nov 26 14:48:47 2018  step: 8700  accuracy: 0.96
Mon Nov 26 14:52:27 2018  step: 8800  accuracy: 0.9625
Mon Nov 26 14:56:06 2018  step: 8900  accuracy: 0.9525
Mon Nov 26 14:59:45 2018  step: 9000  accuracy: 0.9575
Mon Nov 26 15:03:24 2018  step: 9100  accuracy: 0.9675
Mon Nov 26 15:07:02 2018  step: 9200  accuracy: 0.9725
Mon Nov 26 15:10:41 2018  step: 9300  accuracy: 0.9575
Mon Nov 26 15:14:19 2018  step: 9400  accuracy: 0.97
Mon Nov 26 15:17:57 2018  step: 9500  accuracy: 0.9775
Mon Nov 26 15:21:36 2018  step: 9600  accuracy: 0.9775
Mon Nov 26 15:25:14 2018  step: 9700  accuracy: 0.97
Mon Nov 26 15:28:53 2018  step: 9800  accuracy: 0.98
Mon Nov 26 15:32:32 2018  step: 9900  accuracy: 0.9825
Mon Nov 26 15:36:10 2018  step: 10000  accuracy: 0.9775
Mon Nov 26 15:39:48 2018  step: 10100  accuracy: 0.97
Mon Nov 26 15:43:26 2018  step: 10200  accuracy: 0.9575
Mon Nov 26 15:47:05 2018  step: 10300  accuracy: 0.9625
Mon Nov 26 15:50:42 2018  step: 10400  accuracy: 0.9675
Mon Nov 26 15:54:21 2018  step: 10500  accuracy: 0.9725
Mon Nov 26 15:57:59 2018  step: 10600  accuracy: 0.9825
Mon Nov 26 16:01:39 2018  step: 10700  accuracy: 0.9825
Mon Nov 26 16:05:18 2018  step: 10800  accuracy: 0.985
Mon Nov 26 16:08:57 2018  step: 10900  accuracy: 0.9625
Mon Nov 26 16:12:38 2018  step: 11000  accuracy: 0.965
Mon Nov 26 16:16:18 2018  step: 11100  accuracy: 0.97
Mon Nov 26 16:19:58 2018  step: 11200  accuracy: 0.9875
Mon Nov 26 16:23:37 2018  step: 11300  accuracy: 0.98
Mon Nov 26 16:27:17 2018  step: 11400  accuracy: 0.9625
Mon Nov 26 16:30:58 2018  step: 11500  accuracy: 0.975
Mon Nov 26 16:34:39 2018  step: 11600  accuracy: 0.985
Mon Nov 26 16:38:18 2018  step: 11700  accuracy: 0.9775
Mon Nov 26 16:41:58 2018  step: 11800  accuracy: 0.9825
Mon Nov 26 16:45:38 2018  step: 11900  accuracy: 0.98
Mon Nov 26 16:49:18 2018  step: 12000  accuracy: 0.9825
Mon Nov 26 16:52:57 2018  step: 12100  accuracy: 0.985
Mon Nov 26 16:56:36 2018  step: 12200  accuracy: 0.98
Mon Nov 26 17:00:15 2018  step: 12300  accuracy: 0.9725
Mon Nov 26 17:03:54 2018  step: 12400  accuracy: 0.955
Mon Nov 26 17:07:35 2018  step: 12500  accuracy: 0.97
Mon Nov 26 17:11:14 2018  step: 12600  accuracy: 0.9775
Mon Nov 26 17:14:53 2018  step: 12700  accuracy: 0.985
...

captcha_check's Issues

could not convert string to float

Traceback (most recent call last):
File "model_test.py", line 34, in
image = captcha2text('img/2uB3.jpg')
File "model_test.py", line 27, in captcha2text
vector_list = sess.run(predict, feed_dict={x: image_list, keep_prob: 1})
File "/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 889, in run
run_metadata_ptr)
File "/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/tensorflow/python/client/session.py", line 1089, in _run
np_val = np.asarray(subfeed_val, dtype=subfeed_dtype)
File "/Library/Frameworks/Python.framework/Versions/3.6/lib/python3.6/site-packages/numpy/core/numeric.py", line 501, in asarray
return array(a, dtype, copy=False, order=order)
ValueError: could not convert string to float: 'img/2uB3.jpg'

我自己定义的图片路径 报错诶，不用gen_captcha_text_and_image生成的图片
image = captcha2text('img/2uB3.jpg')

图片大小修改后，无法运行。

请问一下，我修改了生成图片的大小为120x40后，运行model_train脚本，却一直卡在一个地方。

[已解决]如果要基于现有模型训练怎么改呢

如果要基于现有模型训练怎么改呢

[已解决]转换成pb模型输出层应该怎么写呢？

我遍历output_graph_def.node，打印每一个node的name，最后一个节点名字是init，但是使用pb模型的时候init层并没有数据，请问输出层应该怎没写呢？

转换脚本

# -*-coding: utf-8 -*-
"""
    参考自：
    @Project: tensorflow_models_nets
    @File   : convert_pb.py
    @Author : panjq
    @E-mail : [email protected]
    @Date   : 2018-08-29 17:46:50
    @info   :
    -通过传入 CKPT 模型的路径得到模型的图和变量数据
    -通过 import_meta_graph 导入模型中的图
    -通过 saver.restore 从模型中恢复图中各个变量的数据
    -通过 graph_util.convert_variables_to_constants 将模型持久化
"""

import tensorflow as tf
from tensorflow.python.framework import graph_util


def freeze_graph(input_checkpoint, output_graph):
    '''
    :param input_checkpoint:
    :param output_graph: PB模型保存路径
    :return:
    '''
    # checkpoint = tf.train.get_checkpoint_state(model_folder) #检查目录下ckpt文件状态是否可用
    # input_checkpoint = checkpoint.model_checkpoint_path #得ckpt文件路径

    # 指定输出的节点名称,该节点名称必须是原模型中存在的节点
    output_node_names = ["init"]
    # output_node_names = [ "YoloV3/pred_sbbox/concat_2", "YoloV3/pred_mbbox/concat_2", "YoloV3/pred_lbbox/concat_2"]
    saver = tf.train.import_meta_graph(input_checkpoint + '.meta', clear_devices=True)

    with tf.Session() as sess:
        saver.restore(sess, input_checkpoint)  # 恢复图并得到数据
        output_graph_def = graph_util.convert_variables_to_constants(  # 模型持久化，将变量值固定
            sess=sess,
            input_graph_def=sess.graph_def,  # 等于:sess.graph_def
            output_node_names=output_node_names)  # 如果有多个输出节点，以逗号隔开

        # fix nodes
        for node in output_graph_def.node:
            if node.op == 'RefSwitch':
                node.op = 'Switch'
                for index in range(len(node.input)):
                    if 'moving_' in node.input[index]:
                        node.input[index] = node.input[index] + '/read'
            elif node.op == 'AssignSub':
                node.op = 'Sub'
                if 'use_locking' in node.attr: del node.attr['use_locking']
            elif node.op == 'AssignAdd':
                node.op = 'Add'
                if 'use_locking' in node.attr: del node.attr['use_locking']
            elif node.op == 'Assign':
                node.op = 'Identity'
                if 'use_locking' in node.attr: del node.attr['use_locking']
                if 'validate_shape' in node.attr: del node.attr['validate_shape']
                if len(node.input) == 2:
                    # input0: ref: Should be from a Variable node. May be uninitialized.
                    # input1: value: The value to be assigned to the variable.
                    node.input[0] = node.input[1]
                    del node.input[1]
        for node in output_graph_def.node:
            print(node.name)




        with tf.gfile.GFile(output_graph, "wb") as f:  # 保存模型
            f.write(output_graph_def.SerializeToString())  # 序列化输出
        print("%d ops in the final graph." % len(output_graph_def.node))  # 得到当前图有几个操作节点

        # for op in sess.graph.get_operations():
        #    print(op.name, op.values())


if __name__ == '__main__':
    # 输入ckpt模型路径
    input_checkpoint = 'model/captcha.model-0'
    # 输出pb模型的路径
    out_pb_path = "model/frozen_model.pb"
    # 调用freeze_graph将ckpt转为pb
    freeze_graph(input_checkpoint, out_pb_path)

调用脚本:

import tensorflow as tf
from tensorflow.python.platform import gfile

from captcha_gen import CAPTCHA_LIST, CAPTCHA_LEN
from util import get_next_batch

sess = tf.Session()
with gfile.FastGFile('model/frozen_model.pb', 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())
    sess.graph.as_default()
    tf.import_graph_def(graph_def, name='')  # 导入计算图

sess.run(tf.global_variables_initializer())


x = tf.placeholder(tf.float32, [None, 160 * 60])
y = tf.placeholder(tf.float32, [None, len(CAPTCHA_LIST) * CAPTCHA_LEN])
keep_prob = tf.placeholder(tf.float32)
batch_x, batch_y = get_next_batch(64)

op = sess.graph.get_tensor_by_name('init:0')

ret = sess.run(op, feed_dict={x: batch_x, y: batch_y, keep_prob: 1.0})
print(ret)

调用报错：

KeyError: "The name 'init:0' refers to a Tensor which does not exist. The operation, 'init', exists but only has 0 outputs."