基于 CNN 的字符识别

本文声明：本文只本人在学习过程中的练习，如有错误之处，欢迎大家不吝赐教

在学习过程中，对标准的打印字符构建模型，并进行训练，字符样本如下图所示：

模型中共用了 3 层卷积和2 个全连接层，其中，第1层卷积核大小为 3*3，输入图像为单通道灰度图像，输出特征为 64 个特征，第2层卷积核大小为5*5，输出特征为32个特征，第3层卷积核大小为5*5，输出特征为 16 个特征，将 16 个特征在第一层全连接层，映射为 128 个特征，第2层全连接层，将上一层的128个特征，映射为10个分类；

其中，每个字符选取了10个样本，并将样本进行归一化大小为128*256大小的图像，进行训练，迭代次数为100；训练过程中，损失值与准确率如下图所示：

从上图可以看书，在训练到20次左右，Loss值趋向于 0,说明是收敛的，而且准确率，也逐步达了100%，因为样本全部为标准字符，所以，识别的准确率达到100%，但也可能存在过拟合；

实现代码如下所示：

# 该程序使用 TensorFlow 对 CNN 进行实现
# 图像预处理部分使用 OpenCV

import os
import cv2
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'        # 防止报出 TensorFlow 中的警告


# 以下代码用于实现卷积网络

def weight_init(shape, name):
    return tf.get_variable(name, shape, initializer=tf.random_normal_initializer(mean=0.0, stddev=0.1))


def bias_init(shape, name):
    return  tf.get_variable(name, shape, initializer=tf.constant_initializer(0.0))


def conv2d(input_data, conv_w):
    return tf.nn.conv2d(input_data, conv_w, strides=[1, 1, 1, 1], padding='VALID')


def max_pool(input_data, size):
    return tf.nn.max_pool(input_data, ksize=[1, size, size, 1], strides=[1, size, size, 1], padding='VALID')


def conv_net(input_data):

    with tf.name_scope('conv1'):
        w_conv1 = weight_init([3, 3, 1, 64], 'conv1_w')  # 卷积核大小是 3*3 输入是 1 通道,输出为 64 通道
        b_conv1 = bias_init([64], 'conv1_b')
        h_conv1 = tf.nn.relu(tf.nn.bias_add(conv2d(input_data, w_conv1), b_conv1))
        h_pool1 = max_pool(h_conv1, 2)

    with tf.name_scope('conv2'):
        w_conv2 = weight_init([5, 5, 64, 32], 'conv2_w')  # 卷积核大小是 5*5 输入是64,输出为 32
        b_conv2 = bias_init([32], 'conv2_b')
        h_conv2 = tf.nn.relu(conv2d(h_pool1, w_conv2) + b_conv2)
        h_pool2 = max_pool(h_conv2, 2)

    with tf.name_scope('conv3'):
        w_conv3 = weight_init([5, 5, 32, 16], 'conv3_w')  # 卷积核大小是 5*5 输入是32,输出为 16
        b_conv3 = bias_init([16], 'conv3_b')
        h_conv3 = tf.nn.relu(conv2d(h_pool2, w_conv3) + b_conv3)
        h_pool3 = max_pool(h_conv3, 2)
        # print(h_pool3)

    with tf.name_scope('fc1'):
        w_fc1 = weight_init([28 * 12 * 16, 128], 'fc1_w')  # 三层卷积后得到的图像大小为 28 * 12
        b_fc1 = bias_init([128], 'fc1_b')
        h_fc1 = tf.nn.relu(tf.matmul(tf.reshape(h_pool3, [-1, 28*12*16]), w_fc1) + b_fc1)


    with tf.name_scope('fc2'):
        w_fc2 = weight_init([128, 10], 'fc2_w')
        b_fc2 = bias_init([10], 'fc2_b')
        h_fc2 = tf.matmul(h_fc1, w_fc2) + b_fc2


    return h_fc2


"""
# 利用 os 模块中的 walk() 函数遍历文件
def load_images(path):
    image_list = []     # 图像列表,用于保存加载到内存中的图像文件
    for dirPath, _, fileNames in os.walk(path):
        for fileName in fileNames:
            if fileName == '001.jpg':
                image_name = os.path.join(dirPath, fileName)
                image = cv2.imread(image_name, cv2.IMREAD_GRAYSCALE)        # 加载为灰度图像
                image_list.append(image)


    return image_list
"""

def load_images():
    img_list = []
    for i in range(10):     # 文件夹
        path = 'F:\Python WorkSpace\OcrByCNN\Characters\%d\\' % i

        for j in range(10): # 文件名
            file_names = '%03d.jpg' % (j+1)
            image = cv2.imread(path + file_names, cv2.IMREAD_GRAYSCALE)
            img_list.append(image)


    return img_list





def get_accuracy(logits, targets):
    batch_prediction = np.argmax(logits, axis=1)
    num_correct = np.sum(np.equal(batch_prediction, targets))
    return 100.* num_correct / batch_prediction.shape[0]


def train_data():
    image_list = load_images()

    batch_size = 100
    np_train_datas = np.empty(shape=(batch_size, 256, 128, 1), dtype='float32')

    for i in range(100):
        np_train_datas[i] = image_list[i][ :, :, np.newaxis]/256  # 归一化



    # train_datas = tf.convert_to_tensor(np_train_datas)
    train_labels_data = np.zeros(shape=(100,), dtype=int)

    for i in range(10):
        train_labels_data[i*10: 10*(i + 1)] = i

    train_labels = tf.one_hot(train_labels_data, 10, on_value=1.0, off_value=0.0)

    """
    sess = tf.Session()
    init = tf.global_variables_initializer()
    sess.run(init)

    print(sess.run(train_labels))

    """
    x_data = tf.placeholder(tf.float32, [batch_size, 256, 128, 1])
    y_target = tf.placeholder(tf.float32, [batch_size, 10])

    model_output = conv_net(x_data)
    # print(predict)
    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels=y_target, logits=model_output))
    # loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=model_output, logits=y_target), name='loss')
    optimizer = tf.train.AdamOptimizer(1e-2).minimize(loss)
    # prediction = tf.nn.softmax(model_output)

    # 计算训练集准确率
    train_correct_prediction = tf.equal(tf.argmax(model_output, 1), tf.argmax(train_labels, 1))
    train_accuracy_1 = tf.reduce_mean(tf.cast(train_correct_prediction, tf.float32))

    init = tf.global_variables_initializer()

    session = tf.Session()
    session.run(init)

    train_loss = []
    train_accuracy = []

    for i in range(100):        # 训练 100 次
        [_loss, _, accuracy] = session.run([loss, optimizer, train_accuracy_1], feed_dict={x_data: np_train_datas, y_target: train_labels.eval(session=session)})
        # accuracy = get_accuracy(_prediction, y_target)
        train_loss.append(_loss)
        train_accuracy.append(accuracy)

        print('第 %d 次迭代:' % i)
        print('loss: %0.10f' % _loss)
        print('perdiction: %0.2f' % accuracy)
         # print('prediction: %f' % _prediction)

    session.close()

    plt.title('train loss')
    plt.plot(range(0, 100), train_loss, 'b-')
    plt.show()
    plt.title('accuracy')
    plt.plot(range(0, 100), train_accuracy, 'k-')
    plt.show()



def main():
    train_data()

if __name__ == "__main__":
    main()