Keras神经网络识别黑白验证码

# print(ord('0'),ord('A'),ord('a')) # ascii码  0-48  A-65  a-97
# print(10+26+26) # 62个维度

import numpy as np

# 62维特征向量的排列按如下内存结构
# 10位数字 + 26位大写字母 + 26位小写字母
# [0,0,0,0,0,0....,0,0]

def char2pos(c:str):
    """
    将一个字符转化为62位特征向量中的位置
    :param c:
    :return:
    """
    c_idx = ord(c)
    # 可能为0-9 a-z A-Z
    # 减去ascii码偏移量，加上前面数字字母的长度
    if c_idx < 65:
        return c_idx - 48 # 0-9
    elif c_idx < 97:
        return c_idx - 65 + 10 # A-Z
    else:
        return c_idx - 97 + 10 + 26 # a-z


# print(char2pos('a'))

def text2vec(content:str):
    """
    将字符串转换为特诊向量
    :param content:
    :return:
    """
    result = np.zeros((62*6,))
    for idx,c in enumerate(content):
        base_idx = idx * 62
        result[base_idx + char2pos(c)] = 1

    return result

import os
import matplotlib.pyplot as plt

def load_img():
    """
    加载所有的训练数据
    :return:
    """
    img_names = os.listdir('./img')

    # 这里仅有32张图片
    train_data = np.zeros((32,50 * 240))
    target_data = np.zeros((32,62 * 6)) # 验证码有6个字符

    for idx,name_str in enumerate(img_names):
        img = plt.imread('./img/{}'.format(name_str))
        # 获取验证码内容，验证码内容用图片名称表示
        img_name = name_str.split('.')[0]

        # 将图片内容和目标分别进行赋值
        train_data[idx] = img.reshape(-1)
        target_data[idx] = text2vec(img_name)

    return train_data,target_data


def vec2text(v:np.array):
    """
    将预测结果转换为字符串
    :param v:
    :return:
    """
    idx_all = v.nonzero()[0]
    result = []

    for idx in idx_all:
        idx_char = idx % 62
        real_char = ''
        if idx_char < 10:
            # 数字
            real_char = idx_char + ord('0')
        elif idx_char < 36:
            real_char = idx_char - 10 + ord('A')
        else:
            real_char = idx_char - 10 - 26 + ord('a')

        result.append(chr(real_char))

    return ''.join(result)

# print(vec2text(text2vec('8uf2wj')))

from keras.models import Sequential,load_model
from keras.layers import Conv2D,MaxPool2D,Flatten,Dense
import tensorflow as tf
from sklearn.metrics import r2_score

train_data,target_data = load_img()
train_data = train_data.reshape(-1,50,240,1).astype(np.float32)
target_data = target_data.astype(np.float32)

model = Sequential()

model.add(Conv2D(16,(3,3),input_shape=(50,240,1),activation=tf.nn.relu,padding='SAME')) # 1-单通道
model.add(MaxPool2D(pool_size=(2,2)))

model.add(Conv2D(8,(3,3),padding='SAME',activation=tf.nn.relu))
model.add(MaxPool2D(pool_size=(2,2)))

model.add(Flatten())
model.add(Dense(512,activation=tf.nn.relu))
model.add(Dense(372,activation=tf.nn.softmax))
# from keras.losses import categorical_crossentropy
model.compile(loss='categorical_crossentropy',optimizer='Adam',metrics=['categorical_crossentropy'])

model.fit(train_data,target_data,epochs=1000)

print('r2_score is ::::',r2_score(target_data,model.predict(train_data)))