卷积神经网络手写数字识别

卷积神经网络手写数字识别（基于TensorFlow），数据：MNIST库（手写体数字库），包含55000张训练图片，每张图片分辨率是28×28，故训练网络输入应该是28×28=784个像素数据。

  • 训练、保存、测试模型

    1
    from PIL import Image
    2
    import matplotlib.pyplot as plt
    3
    import tensorflow as tf
    4
    from tensorflow.examples.tutorials.mnist import input_data
    5
    6
    # 导入 MNIST 数据集
    7
    mnist = input_data.read_data_sets('./MNIST_data/', one_hot=True)
    8
    9
    # 训练参数
    10
    learning_rate = 0.001
    11
    num_steps = 5000
    12
    batch_size = 128
    13
    display_step = 10
    14
    15
    # 神经网络网络参数
    16
    num_input = 784  # MNIST 输入数据规格(img shape: 28*28)
    17
    num_classes = 10  # MNIST 类数(0-9 digits)
    18
    dropout = 0.75  # Dropout, 保留神经元的概率
    19
    20
    # TensorFlow输入
    21
    X = tf.placeholder(tf.float32, [None, num_input])
    22
    Y = tf.placeholder(tf.float32, [None, num_classes])
    23
    keep_prob = tf.placeholder(tf.float32)  # dropout (保留概率)
    24
    25
    26
    # 卷积
    27
    def conv2d(x, W, b, strides=1):
    28
        x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    29
        x = tf.nn.bias_add(x, b)
    30
        return tf.nn.relu(x)
    31
    32
    33
    # 池化
    34
    def maxpool2d(x, k=2):
    35
        return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME')
    36
    37
    38
    # 创建卷积神经网络模型
    39
    def conv_net(x, weights, biases, dropout):
    40
        # MNIST data input is a 1-D vector of 784 features (28*28 pixels)
    41
        # Reshape to match picture format [Height x Width x Channel]
    42
        # Tensor input become 4-D: [Batch Size, Height, Width, Channel]
    43
        x = tf.reshape(x, shape=[-1, 28, 28, 1])
    44
    45
        # 第一层卷积和池化
    46
        conv1 = conv2d(x, weights['wc1'], biases['bc1'])
    47
        conv1 = maxpool2d(conv1, k=2)
    48
    49
        # 第二层卷积和池化
    50
        conv2 = conv2d(conv1, weights['wc2'], biases['bc2'])
    51
        conv2 = maxpool2d(conv2, k=2)
    52
    53
        # 全连接层
    54
        # Reshape conv2 output to fit fully connected layer input
    55
        fc1 = tf.reshape(conv2, [-1, weights['wd1'].get_shape().as_list()[0]])
    56
        fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])
    57
        fc1 = tf.nn.relu(fc1)
    58
        fc1 = tf.nn.dropout(fc1, rate=1-dropout)  # Apply Dropout
    59
    60
        # 输出层
    61
        out = tf.add(tf.matmul(fc1, weights['out']), biases['out'])
    62
        return out
    63
    64
      
    65
    # 每层的 weight 和 bias
    66
    weights = {
    67
        # 5x5 conv, 1 input, 32 outputs
    68
        'wc1': tf.Variable(tf.random_normal([5, 5, 1, 32])),
    69
        # 5x5 conv, 32 inputs, 64 outputs
    70
        'wc2': tf.Variable(tf.random_normal([5, 5, 32, 64])),
    71
        # fully connected, 7*7*64 inputs, 1024 outputs
    72
        'wd1': tf.Variable(tf.random_normal([7*7*64, 1024])),
    73
        # 1024 inputs, 10 outputs (class prediction)
    74
        'out': tf.Variable(tf.random_normal([1024, num_classes]))
    75
    }
    76
    biases = {
    77
        'bc1': tf.Variable(tf.random_normal([32])),
    78
        'bc2': tf.Variable(tf.random_normal([64])),
    79
        'bd1': tf.Variable(tf.random_normal([1024])),
    80
        'out': tf.Variable(tf.random_normal([num_classes]))
    81
    }
    82
    83
    # 建立模型
    84
    logits = conv_net(X, weights, biases, keep_prob)
    85
    prediction = tf.nn.softmax(logits)
    86
    87
    # 损失函数与优化器
    88
    loss_op = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=Y))
    89
    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
    90
    train_op = optimizer.minimize(loss_op)
    91
    92
    # 评估模型
    93
    correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y, 1))
    94
    accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
    95
    96
    # 初始化变量
    97
    init = tf.global_variables_initializer()
    98
    99
    # 定义saver: 保存模型
    100
    saver = tf.train.Saver()
    101
    102
    103
    # 训练模型
    104
    def train_model():
    105
        with tf.Session() as sess:
    106
            sess.run(init)
    107
    108
            for step in range(1, num_steps+1):
    109
                batch_x, batch_y = mnist.train.next_batch(batch_size)
    110
                # 反向传播优化
    111
                sess.run(train_op, feed_dict={X: batch_x, Y: batch_y, keep_prob: dropout})
    112
                if step % display_step == 0 or step == 1:
    113
                    # Calculate batch loss and accuracy
    114
                    loss, acc = sess.run([loss_op, accuracy], feed_dict={X: batch_x, Y: batch_y, keep_prob: 1.0})
    115
                    print("Step " + str(step) + ", Minibatch Loss= " + "{:.4f}".format(loss) + ", Training Accuracy= " + "{:.3f}".format(acc))
    116
            saver.save(sess, 'CNN_handwritten_digits_model')  # 储存模型
    117
            print("Optimization Finished!")
    118
    119
            # 计算测试精度(256 MNIST test images)
    120
            print("Testing Accuracy:", sess.run(accuracy, feed_dict={X: mnist.test.images[:256], Y: mnist.test.labels[:256], keep_prob: 1.0}))
    121
    122
    123
    # 图片预处理
    124
    def image_prepare(picture):
    125
        image = Image.open(picture)  # 读取图片(28*28)
    126
        plt.imshow(image)  # 显示需要识别的图片
    127
        plt.show()
    128
        image = image.convert('L')
    129
        image = image.resize((28, 28), Image.ANTIALIAS)
    130
        image_data = list(image.getdata())
    131
        image_data = [0 if x < 50 else 1 for x in image_data]  # 如果黑底白字
    132
        # image_data = [1 if x < 50 else 0 for x in image_data]  # 如果白底黑字
    133
        # cnt = 0
    134
        # for i in image_data:
    135
        #     print(i, end=' ')
    136
        #     cnt += 1
    137
        #     if cnt % 28 == 0:
    138
        #         print(end='\n')
    139
        return image_data
    140
    141
    142
    # 待识别数字图片
    143
    image = image_prepare('./digits/62.png')
    144
    145
    # 训练模型
    146
    # train_model()
    147
    148
    # 测试模型
    149
    with tf.Session() as sess:
    150
        saver = tf.train.import_meta_graph('CNN_handwritten_digits_model.meta')
    151
        saver.restore(sess, tf.train.latest_checkpoint('./'))
    152
    153
        prediction = tf.argmax(prediction, 1)
    154
        predint = prediction.eval(feed_dict={X: [image], keep_prob: 1.0}, session=sess)
    155
    156
        print('识别结果:', end=' ')
    157
        print(predint[0])