HIT 模式识别 手写汉字分类 Python实现

训练集数据 TrainSamples-400.csv，含 100 个不同汉字，每个汉字 400 个实例，每个实例均为 64*64 的二值图像；
训练集标注TrainSamples-400.csv，为 40000 个 0 到 99 间的整数，表示训练集中每个实例所属汉字类别；
测试集数据 TestSamples-300.csv，为 30000 个实例，每个实例格式同训练集。
要求标注测试集，输出 Result.csv。

CNN 网络结构：

代码：

import numpy as np
import pandas as pd
from tensorflow.keras.utils import to_categorical
from tensorflow.keras import models, layers

def train():
    data = pd.read_csv("TrainSamples-400.csv", header=None)
    train_image = data.to_numpy()
    data = pd.read_csv("TrainLabels-400.csv", header=None)
    train_label = data.to_numpy()
    train_label = to_categorical(train_label)
    network = models.Sequential()
    network.add(layers.Input(shape = (64, 64, 1)))
    network.add(layers.Conv2D(64, (5, 5), activation = 'relu'))
    network.add(layers.MaxPooling2D((2, 2)))
    network.add(layers.Conv2D(96, (3, 3), activation = 'relu'))
    network.add(layers.MaxPooling2D((2, 2)))
    network.add(layers.Conv2D(48, (3, 3), activation = 'relu'))
    network.add(layers.Flatten())
    network.add(layers.Dense(768, activation = 'relu'))
    network.add(layers.Dense(100, activation = 'softmax'))
    network.summary()
    network.compile(optimizer = 'rmsprop', loss = 'categorical_crossentropy', metrics = ['accuracy'])
    network.fit(train_image.reshape(40000, 64, 64, 1), train_label, epochs = 5, batch_size = 64, validation_split = 0.1, validation_freq = 1)
    network.save('saved_model/my_model')
    
def test():
    data = pd.read_csv("TestSamples-300.csv", header = None)
    test_image = data.to_numpy()
    network = models.load_model('saved_model/my_model')
    network.summary()
    test_label = network.predict(test_image.reshape(30000, 64, 64, 1))
    test_label = np.array([np.argmax(i) for i in test_label])
    pd.DataFrame(test_label).to_csv('Result.csv', header = None, index = False)

if __name__ == '__main__':
    train()
    test()

结果：