基于python3.6+tensorflow2.2的手写体识别案例

unzip_save.py

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'  # 不显示等级2以下的提示信息
import zipfile
import csv
import numpy as np
import tensorflow as tf
from tensorflow.keras.preprocessing.image import ImageDataGenerator


# 解压
local_zip1 = 'E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_train.zip' # 数据集压缩包路径
zip_ref1 = zipfile.ZipFile(local_zip1, 'r') # 打开压缩包，以读取方式
zip_ref1.extractall('E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_train/') # 解压到以下路径
zip_ref1.close()

local_zip2 = 'E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_test.zip' # 数据集压缩包路径
zip_ref2 = zipfile.ZipFile(local_zip2, 'r') # 打开压缩包，以读取方式
zip_ref2.extractall('E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_test/') # 解压到以下路径
zip_ref2.close()

# # 从google.colab数据库加载文件
# from google.colab import files
#
#
# uploaded = files.upload()

def get_data(filename):
    with open(filename) as training_file:
        csv_reader = csv.reader(training_file, delimiter= ',')
        # csv.reader(file, delimiter=',', quotechar='"')
        # delimiter是分隔符，quotechar是引用符，当一段话中出现分隔符的时候，用引用符将这句话括起来，就能排除歧义。
        first_line = True
        temp_images = []
        temp_labels = []
        for row in csv_reader:
            if first_line:
                # print("Ignoring first line")
                first_line = False # 让for循环跳过第一行数据
            else:
                temp_labels.append(row[0])
                image_data = row[1:785]
                image_data_as_array = np.array_split(image_data, 28) # 将一维ndarray格式转换成size为28*28的ndarray格式
                temp_images.append(image_data_as_array) # 将一维数组添加到list中
        images = np.array(temp_images).astype('float') # 将list格式转换为ndarray三维格式
        labels = np.array(temp_labels).astype('float')
    return images, labels

training_images, training_labels = get_data('E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_train/sign_mnist_train.csv')
testing_images, testing_labels = get_data('E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_test/sign_mnist_test.csv')

# 保存处理后的数据
np.save("E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_train/training_images.npy",training_images)
np.save("E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_train/training_labels.npy",training_labels)

np.save("E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_test/testing_images.npy",testing_images)
np.save("E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_test/testing_labels.npy",testing_labels)

print(training_images.shape)
print(training_labels.shape)
print(testing_images.shape)
print(testing_labels.shape)

---

model_training_fit.py

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'  # 不显示等级2以下的提示信息
# import shutil
# import random
import tensorflow as tf
# from tensorflow import keras
from tensorflow.keras.preprocessing.image import ImageDataGenerator
# from shutil import copyfile
# import matplotlib.image as mpimg
import matplotlib.pyplot as plt
import numpy as np


# 数据加载
training_images = np.load("E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_train/training_images.npy")
training_labels = np.load("E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_train/training_labels.npy")

testing_images = np.load("E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_test/testing_images.npy")
testing_labels = np.load("E:/Python/pythonProject_1/sign_language_mnist/tmp/sign_mnist_test/testing_labels.npy")

# 数据预处理
training_images = np.expand_dims(training_images, axis=3) # 将三维数组扩展到4维
testing_images = np.expand_dims(testing_images, axis=3)

training_datagen = ImageDataGenerator(
    # 数据增强
    rescale=1. / 255,
    rotation_range=40, # 旋转范围
    width_shift_range=0.2, # 宽平移
    height_shift_range=0.2,# 高平移
    shear_range=0.2, # 剪切
    zoom_range=0.2, # 缩放
    horizontal_flip=True,
    fill_mode='nearest'
)

validation_datagen = ImageDataGenerator(
    rescale=1. / 255
)

print(training_images.shape)
print(testing_images.shape)

#======== 模型构建 =========
model = tf.keras.models.Sequential([
    tf.keras.layers.Conv2D(64, (3, 3), activation = 'relu', input_shape = (28, 28, 1)), # 输入参数：过滤器数量，过滤器尺寸，激活函数：relu， 输入图像尺寸
    tf.keras.layers.MaxPooling2D(2, 2), # 池化：增强特征
    tf.keras.layers.Conv2D(64, (3, 3), activation = 'relu'), # 输入参数：过滤器数量、过滤器尺寸、激活函数：relu
    tf.keras.layers.MaxPooling2D(2, 2),
    tf.keras.layers.Flatten(), # 输入层
    tf.keras.layers.Dense(128, activation = 'relu'), # 全连接隐层 神经元数量：128 ，激活函数：relu
    tf.keras.layers.Dense(26, activation = 'softmax') # 英文字母分类 26 ，阿拉伯数字分类 10  输出用的是softmax 概率化函数 使得所有输出加起来为1 0-1之间
])

#======== 模型参数编译 =========
model.compile(
    optimizer = 'adam',
    loss = 'sparse_categorical_crossentropy', # 损失函数： 稀疏的交叉熵 binary_crossentropy
    metrics = ['accuracy']
)

#======== 模型训练 =========
# Note that this may take some time.
history = model.fit_generator(
    training_datagen.flow(training_images, training_labels,batch_size=32),
    steps_per_epoch=len(training_images) / 32,
    epochs = 15,
    validation_data=validation_datagen.flow(testing_images, testing_labels,batch_size=32),
    validation_steps=len(testing_images) / 32
)

model.save('E:/Python/pythonProject_1/sign_language_mnist/model.h5') # model 保存

model.evaluate(testing_images, testing_labels )# 数据(data)和金标准(label),然后将预测结果与金标准相比较,得到两者误差并输出.
#-----------------------------------------------------------
# Retrieve a list of list result on training and test data
# set for each training epoch
#-----------------------------------------------------------
acc = history.history['accuracy']
val_acc = history.history['val_accuracy']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(len(acc)) # Get number of epochs

#-----------------------------------------------------------
# Plot training and validation accuracy per epoch
#-----------------------------------------------------------
plt.plot(epochs, acc, 'r', label = "tra_acc")
plt.plot(epochs ,val_acc, 'b', label = "val_acc")
plt.title("training and validation accuracy")
plt.legend()
plt.grid(ls='--')  # 生成网格
plt.show()
# 曲线呈直线是因为epochs/轮次太少
#-----------------------------------------------------------
# Plot training and validation loss per epoch
#-----------------------------------------------------------
plt.plot(epochs, loss, 'r', label = "train_loss")
plt.plot(epochs ,val_loss, 'b', label = "val_loss")
plt.title("training and validation loss")
plt.legend()
plt.grid(ls='--')  # 生成网格
plt.show()
# 曲线呈直线是因为epochs/轮次太少

---

测试结果：

Epoch 15/15
858/857 [==============================] - 25s 29ms/step - loss: 0.7643 - accuracy: 0.7437 - val_loss: 0.3325 - val_accuracy: 0.8799

 

---

评估结果：

225/225 [==============================] - 1s 6ms/step - loss: 240.8037 - accuracy: 0.6163

---

predict.py

# -*- coding: UTF-8 -*-
import numpy as np
import matplotlib.image as mpig
from tensorflow import keras
from PIL import Image

model = keras.models.load_model('E:/Python/pythonProject_1/sign_language_mnist/model.h5')

im = Image.open('E:/Python/pythonProject_1/sign_language_mnist/25.png')
im_resized = im.resize((28, 28))

img = im_resized.convert('L')
img.save('E:/Python/pythonProject_1/sign_language_mnist/25.png')

img = mpig.imread('E:/Python/pythonProject_1/sign_language_mnist/25.png')
img = img * 255
img = img.astype(np.float64)
images = np.expand_dims(img, axis=0) # 将三维数组扩展到3维
images = np.expand_dims(images, axis=3) # 将三维数组扩展到4维
classes = model.predict(images, batch_size=26)


result = np.argmax(classes, axis=1)
print(result)

24号美式字母手势，代表字母“Y”

24种美式字母手势

预处理后：

 

Result:

>>> print(result)
[24]

---

备注： 

数据库链接：https://www.kaggle.com/datamunge/sign-language-mnist