昨天通过keras构建简单神经网络实现手写数字识别，结果在最后进行我们自己的手写数字识别的时候，准确率堪忧，只有60%。今天通过卷积神经网络来实现手写数字识别。
构建卷积神经网络和简单神经网络思路类似，只不过这里加入了卷积、池化等概念，网络结构复杂了一些，但是整体的思路没有变化，加载数据集，数据集修改，搭建网络模型，编译模型，训练模型，保存模型，利用模型预测。
这里还是给出两个例子，一个是构建网络，最后保存训练好的网络模型，一个是通过加载保存的网络模型预测我们自己的手写数字图片。
import keras
import numpy as np
import tensorflow as tf
from keras.models import Sequential
from keras.layers import Dense, Activation, Dropout, Conv2D, Flatten, MaxPool2D
from tensorflow.keras import datasets, utils
# 数据处理
(x_train, y_train), (x_test, y_test) = datasets.mnist.load_data()
x_train = x_train.reshape(x_train.shape[0], x_train.shape[1], x_train.shape[1], 1)
x_train = x_train.astype('float32') / 255
x_test = x_test.reshape(x_test.shape[0], x_test.shape[1], x_test.shape[1], 1)
x_test = x_test.astype('float32') / 255
y_train = utils.to_categorical(y_train, num_classes=10)
y_test = utils.to_categorical(y_test, num_classes=10)
# 构建模型
model = Sequential()
model.add(Conv2D(filters=16, kernel_size=(3, 3), padding='same', activation="relu", input_shape=(28, 28, 1)))
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Conv2D(filters=36, kernel_size=(3, 3), padding='same', activation="relu"))
model.add(MaxPool2D(pool_size=(2, 2)))
model.add(Dropout(0.2))
model.add(Flatten())
model.add(Dense(128, activation="relu"))
model.add(Dropout(0.25))
model.add(Dense(10, activation="softmax"))
# 编译
model.compile(loss="categorical_crossentropy", optimizer="adam", metrics=["accuracy"])
model.summary()
# 训练
model.fit(x_train, y_train, epochs=5, batch_size=128, validation_data=(x_test, y_test))
# 保存模型
model.save("mnist.h5")
训练模型，打印信息如下：
Model: "sequential"
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d (Conv2D) (None, 28, 28, 16) 160
max_pooling2d (MaxPooling2D (None, 14, 14, 16) 0
)
conv2d_1 (Conv2D) (None, 14, 14, 36) 5220
max_pooling2d_1 (MaxPooling (None, 7, 7, 36) 0
2D)
dropout (Dropout) (None, 7, 7, 36) 0
flatten (Flatten) (None, 1764) 0
dense (Dense) (None, 128) 225920
dropout_1 (Dropout) (None, 128) 0
dense_1 (Dense) (None, 10) 1290
=================================================================
Total params: 232,590
Trainable params: 232,590
Non-trainable params: 0
_________________________________________________________________
Epoch 1/5
2023-08-28 16:03:54.677314: I tensorflow/stream_executor/cuda/cuda_dnn.cc:368] Loaded cuDNN version 8800
469/469 [==============================] - 10s 17ms/step - loss: 0.2842 - accuracy: 0.9123 - val_loss: 0.0628 - val_accuracy: 0.9798
Epoch 2/5
469/469 [==============================] - 7s 16ms/step - loss: 0.0836 - accuracy: 0.9743 - val_loss: 0.0473 - val_accuracy: 0.9841
Epoch 3/5
469/469 [==============================] - 7s 16ms/step - loss: 0.0627 - accuracy: 0.9801 - val_loss: 0.0325 - val_accuracy: 0.9886
Epoch 4/5
469/469 [==============================] - 7s 15ms/step - loss: 0.0497 - accuracy: 0.9844 - val_loss: 0.0346 - val_accuracy: 0.9882
Epoch 5/5
469/469 [==============================] - 7s 15ms/step - loss: 0.0422 - accuracy: 0.9867 - val_loss: 0.0298 - val_accuracy: 0.9898准确率最后，到达了98.5%以上。
用模型预测
import keras
import numpy as np
import cv2
from keras.models import load_model
model = load_model("mnist.h5")
def predict(img_path):
img = cv2.imread(img_path, 0)
img = img.reshape(28, 28).astype("float32") / 255 # 0 1
img = img.reshape(1, 28, 28, 1) # 28 * 28 -> (1,28,28,1)
label = model.predict(img)
label = np.argmax(label, axis=1)
print('{} -> {}'.format(img_path, label[0]))
if __name__ == '__main__':
for _ in range(10):
predict("number_images/b_{}.png".format(_))
数字图片如下：
图片放在项目目录number_images中。
预测结果打印：
感觉就是不一样，准确率从60%提升到了90%。虽然没有达到100%，但是已经很好了。
对比之前的代码，改动很小，主要是网络输入的时候，数据形状发生了改变，简单神经网络需要的是（784，*）结构，卷积神经网络需要的是（1，28，28，1）的结构， 在数据处理上做了调整，另一个不一样的地方就是网络模型在添加的时候，之前就是简单的两层网络，卷积神经网络复杂了很多。
