机器学习 实验三 手写汉字识别

机器学习 实验三 手写汉字识别

一、实验环境

PC机，Python

二、代码

一、使用神经网络

#%%
import pandas as pd
import tensorflow as tf
import matplotlib.pyplot as plt
import os,PIL,pathlib
import numpy as np
import warnings
from tensorflow import keras
import cv2

#%%
from PIL import Image
img = Image.open(r'data\train\0\input_1_1_1.jpg', 'r')
img

#%%

import os
def file_name(file_dir):
    L=[]
    for root, dirs, files in os.walk(file_dir):
        for file in files:
            if os.path.splitext(file)[1] == '.jpg':
                L.append(os.path.join(root, file))
    L.sort()
    return L

#%%

train_image_paths=file_name('data\\train')
train_path_ds = tf.data.Dataset.from_tensor_slices(train_image_paths)
train_path_ds

#%%

train_image_paths

#%%

# pre_image_paths=file_name('data\\test')
# pre_path_ds = tf.data.Dataset.from_tensor_slices(pre_image_paths)
# pre_path_ds

#%%

# pre_image_paths

#%%

pre_image_paths_sort=[]
dirname='data\\test\\'
for i in range(3000):
    pre_image_paths_sort.append(dirname+str(i)+'.jpg')

#%%

pre_image_paths_sort

#%%

for i in range(3000):
    if(pre_image_paths_sort[i]=='data\\test\\3.jpg'):
        print(i)

#%%

pre_path_ds = tf.data.Dataset.from_tensor_slices(pre_image_paths_sort)
pre_path_ds

#%%

n2=len(pre_path_ds)
n2

#%%

train_image_label=[]
for i in train_image_paths:
    k=i[-6:-4]
    if(k[0]=='_'):
        k=k[-1]
    k=int(k)-1
    train_image_label.append(k)

#将标签切片
train_label_ds = tf.data.Dataset.from_tensor_slices(train_image_label)

#%%

train_image_label

#%%

def preprocess_image(image):
    image = tf.image.decode_jpeg(image,channels = 3)
    image = tf.image.resize(image,[64,64])
    return image / 255.0
def load_and_preprocess_image(path):
    image = tf.io.read_file(path)
    return preprocess_image(image)
#根据路径读取图片并进行预处理
train_image_ds = train_path_ds.map(load_and_preprocess_image,num_parallel_calls=tf.data.experimental.AUTOTUNE)
pre_image_ds = pre_path_ds.map(load_and_preprocess_image,num_parallel_calls=tf.data.experimental.AUTOTUNE)

#%%

train_image_ds

#%%

pre_image_ds

#%%

image_label_ds = tf.data.Dataset.zip((train_image_ds,train_label_ds))

#%%

image_label_ds

#%%

num=0
for i in range(20):
    plt.subplot(4, 5, i + 1)
    num +=1
    plt.xticks([])
    plt.yticks([])
    plt.grid(False)
    t=i+2000
    # 显示图片
    images = plt.imread(train_image_paths[t])
    plt.imshow(images)
    # 显示标签
    plt.xlabel(train_image_label[t])

plt.show()

#%%

n=len(image_label_ds)
n

#%%

image_label_ds = image_label_ds.shuffle(n)

#%%

image_label_ds

#%%

num=0
for i in range(20):
    plt.subplot(4, 5, i + 1)
    num +=1
    plt.xticks([])
    plt.yticks([])
    plt.grid(False)
    t=i+5000
    # 显示图片
    images = plt.imread(train_image_paths[t])
    plt.imshow(images)
    # 显示标签
    plt.xlabel(train_image_label[t])

plt.show()

#%%

test_count=int(n*0.2)
train_count=n-test_count
train_ds = image_label_ds.take(train_count).shuffle(test_count)
test_ds = image_label_ds.skip(train_count).shuffle(test_count)

#%%

train_ds=image_label_ds.shuffle(n)

#%%

train_ds

#%%

pre_ds=pre_image_ds

#%%

# for i in train_ds:
#     print(i)

#%%

height = 64
width = 64
batch_size = 128
epochs = 50

#%%

train_ds = train_ds.batch(batch_size)#设置batch_size
train_ds = train_ds.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
test_ds = test_ds.batch(batch_size)
test_ds = test_ds.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
pre_ds=pre_ds.batch(batch_size)
pre_ds = pre_ds.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)

#%%

plt.figure(figsize=(8, 8))
for images, labels in train_ds.take(1):
    # print(images.shape)
    for i in range(12):
        ax = plt.subplot(4, 3, i + 1)
        plt.imshow(images[i])
        plt.title(labels[i].numpy())  # 使用.numpy()将张量转换为 NumPy 数组
        plt.axis("off")
    break
plt.show()

#%%

train_ds

#%%

# for i in train_ds:
#     print(i)

#%%

test_ds

#%%

pre_ds

#%%

for i in pre_ds:
    print(i)

#%%

model = tf.keras.Sequential([
    tf.keras.layers.Conv2D(filters=32,kernel_size=(3,3),padding="same",activation="relu",input_shape=[64, 64, 3]),
    tf.keras.layers.MaxPooling2D((2,2)),
    tf.keras.layers.Conv2D(filters=64,kernel_size=(3,3),padding="same",activation="relu"),
    tf.keras.layers.MaxPooling2D((2,2)),
    tf.keras.layers.Conv2D(filters=64,kernel_size=(3,3),padding="same",activation="relu"),
    tf.keras.layers.MaxPooling2D((2,2)),
    tf.keras.layers.Flatten(),
    tf.keras.layers.Dense(64, activation="relu"),
    tf.keras.layers.Dense(15, activation="softmax")
])

model.compile(optimizer="adam",
                loss='sparse_categorical_crossentropy',
                metrics=['accuracy'])
model.summary()
history = model.fit(
    train_ds,
    validation_data=test_ds,
    epochs = epochs
)

#%%

model.save('modelbest2.h5')

#%%

train_ds

#%%

test_ds

#%%

modelbest=keras.models.load_model('modelbest2.h5')

#%%

modelbest

#%%

pre=modelbest.predict(pre_ds)
pre[0]

#%%

result=[]
for i in pre:
    maxpre=0
    maxpreindex=0
    for j in range(len(i)):
        if(i[j]>maxpre):
            maxpre=i[j]
            maxpreindex=j
    t=maxpreindex
    if(maxpreindex==11):
        t=100
    elif(maxpreindex==12):
        t=1000
    elif(maxpreindex==13):
        t=10000
    elif(maxpreindex==14):
        t=100000000
    result.append(t)

#%%

result

#%%

plt.figure(figsize=(10, 10))
for images in pre_ds:
    for i in range(20):
        ax = plt.subplot(4, 5, i + 1)
        plt.imshow(images[i])
#         mo="but"
#         s=str(labels[i].numpy())+mo+str(result[i])
        plt.title(result[i])
#         plt.title(pre_image_paths_sort[i][-8:-3])
        plt.axis("off")
    break
plt.show()

#%%

a = [] 
b=[]
for line in range(3000): 
    a.append(str(line)+'.jpg') 
    b.append(result[line])
a = pd.DataFrame(a)
b= pd.DataFrame(b)
df=pd.concat([a,b],axis=1)
df.to_csv('submission', sep='\t', header=None, index=False)

二、使用MLP

#%%

import pandas as pd
import tensorflow as tf
import matplotlib.pyplot as plt
import os,PIL,pathlib
import numpy as np
import warnings
from tensorflow import keras
import cv2
from sklearn.neural_network import MLPClassifier
# train_data = pd.read_csv('/ilab/datasets/local/chinese_num')
# train_data

#%%

from PIL import Image
img = Image.open(r'data\train\0\input_1_1_1.jpg', 'r')
img

#%%

import os
def file_name(file_dir):
    L=[]
    for root, dirs, files in os.walk(file_dir):
        for file in files:
            if os.path.splitext(file)[1] == '.jpg':
                L.append(os.path.join(root, file))
    L.sort()
    return L

#%%

train_image_paths=file_name('data\\train')
train_path_ds = tf.data.Dataset.from_tensor_slices(train_image_paths)
train_path_ds

#%%

pre_image_paths_sort=[]
dirname='data\\test\\'
for i in range(3000):
    pre_image_paths_sort.append(dirname+str(i)+'.jpg')

#%%

for i in range(3000):
    if(pre_image_paths_sort[i]=='data\\test\\3.jpg'):
        print(i)

#%%

pre_path_ds = tf.data.Dataset.from_tensor_slices(pre_image_paths_sort)
pre_path_ds

#%%

train_image_label=[]
for i in train_image_paths:
    k=i[-6:-4]
    if(k[0]=='_'):
        k=k[-1]
    k=int(k)-1
    if(k==11):
        k==100
    elif(k==12):
        k=1000
    elif(k==13):
        k=10000
    elif(k==14):
        k=100000000
    train_image_label.append(k)

#将标签切片
train_label_ds = tf.data.Dataset.from_tensor_slices(train_image_label)

#%%

train_image_label

#%%

def img2vec(fname):
    '''将jpg等格式的图片转为向量'''
    im = Image.open(fname).convert('L')
    im = im.resize((128,128))
    tmp = np.array(im)
    vec = tmp.ravel()/255.0
    return vec

#%%

train_image_data=[]
for i in train_image_paths:
    train_image_data.append(img2vec(i))

#%%

train_image_data

#%%

s_data=np.array(train_image_data)
s_label=np.array(train_image_label)

#%%

index=np.arange(len(s_data))
np.random.shuffle(index)

#%%

s_data=s_data[index]

#%%

s_label=s_label[index]

#%%

pre_image_data=[]
for i in pre_image_paths_sort:
    pre_image_data.append(img2vec(i))

#%%

num=0
for i in range(20):
    plt.subplot(4, 5, i + 1)
    num +=1
    plt.xticks([])
    plt.yticks([])
    plt.grid(False)
    t=i
    # 显示图片
#     images = plt.imread(train_image_data[t])
    plt.imshow(s_data[t].reshape(128,128))
    # 显示标签
    plt.xlabel(s_label[t])

plt.show()

#%%
#尝试lbfgs、sgd、adam三种优化器

# lbfgs = MLPClassifier(solver = 'lbfgs', hidden_layer_sizes = [100,100,100,100,100], activation = 'relu', 
#                       alpha = 1e-4, random_state = 100, verbose = 1)

#%%

sgd = MLPClassifier(solver = 'sgd', hidden_layer_sizes = [300,300,300,300,300], activation = 'relu', 
                    alpha = 1e-4, random_state = 100, verbose = 1, learning_rate_init = 0.1)

#%%

adam = MLPClassifier(solver = 'adam', hidden_layer_sizes = [100,100,100,100], activation = 'relu', 
                      alpha = 1e-4, random_state = 100, verbose = 1)

#%%

from sklearn.model_selection import train_test_split
x_train, x_test, y_train, y_test = train_test_split(s_data, s_label, random_state=1, train_size=0.8)

#%%

# lbfgs.fit(x_train, y_train)

#%%

sgd.fit(x_train, y_train)

#%%

adam.fit(x_train,y_train)

#%%

# lbfgs_predict = lbfgs.predict(x_test)

#%%

sgd_predict = sgd.predict(x_test)

#%%

# adam_predict = adam.predict(x_test)

#%%

# print("lbfgs在训练集准确度: %f" % lbfgs.score(x_train, y_train))
# print("lbfgs在测试集准确度: %f" % lbfgs.score(x_test, y_test))
print("sgd在训练集准确度: %f" % sgd.score(x_train, y_train))
print("sgd在验证集准确度: %f" % sgd.score(x_test, y_test))
print("adam在训练集准确度: %f" % adam.score(x_train, y_train))
print("adam在测试集准确度: %f" % adam.score(x_test, y_test))

#%%

# pre=lbfgs.predict(pre_image_data)

#%%

# pre

#%%

pre2=sgd.predict(pre_image_data)

#%%

# pre2

#%%

# pre3=adam.predict(pre_image_data)
# pre3

#%%

result=[]
for i in pre2:
    result.append(i)

#%%

data=pd.read_csv('submission.csv')

#%%

data

#%%

result

#%%

a = [] 
b=[]
for line in range(3000): 
    a.append(str(line)+'.jpg') 
    b.append(result[line])
a = pd.DataFrame(a)
b= pd.DataFrame(b)
df=pd.concat([a,b],axis=1)
df.to_csv('submissionmlp2.csv', sep='\t', header=None, index=False)

二、实验结果与分析

1、猎豹平台提交结果：

2、对于手写汉字的识别可以使用很多种方法，文中利用的是tensorflow库，大家可以试试pytorch。在MLP中的lbfgs、sgd、adam三种优化器，他们所展现的效果区别还是挺大的，大家可以看看三者的具体区别。（链接我就不放了，大家自由学习）对于多层感知机参数的调整经历了不少的尝试，大家在使用文章代码时可以自主调整，希望能有不错的效果提升。