导入依赖库

import os,cv2,random
import zipfile
import numpy as np
from PIL import Image,ImageQt,ImageDraw,ImageFont
#get_ipython().run_line_magic('matplotlib', 'inline')
import matplotlib
import matplotlib.pyplot as plt
from keras.utils import np_utils
import picklefrom keras.layers import Input,Dense,Dense,Conv2D,Dropout,MaxPooling2D,Flatten,Activation,AveragePooling2D
from keras.optimizers import Adam,SGD,RMSprop,Adadelta
from keras.callbacks import Callback,EarlyStopping,ReduceLROnPlateau
from keras.preprocessing.image import ImageDataGenerator
from keras.applications.vgg16 import VGG16
from keras.applications.resnet50 import ResNet50
from keras.models import Model,Sequential,load_model
from keras.utils import np_utilsfrom sklearn.metrics import accuracy_score

数据预处理

文件准备

数据集链接，百度网盘，9tg6
数据集内存在黑底白字和白底黑字

train_name = 'train-8000-clean.zip'
test_name = 'test1--1200-clean.zip'
train = zipfile.ZipFile(train_name)
test= zipfile.ZipFile(test_name)train_path = './tain_data'
test_path = './test_data'if not os.path.exists(train_path):os.makedirs(train_path)for name in train.namelist():train.extract(name,train_path)
train.close() 
if not os.path.exists(test_path):os.makedirs(test_path)for name in test.namelist():test.extract(name,test_path)
test.close()

图片读取与保存

尺寸28*28，灰度图，使用时按照0-1二值化

ROWS = 28
COLS = 28
CHANNELS = 1
def process_img(data_path):data_list = []data_label = []for i in os.listdir(data_path):data_list.append(i)count = len(data_list)random.shuffle(data_list)#打乱图片名称data = np.ndarray((count, ROWS, COLS), dtype=np.uint8)n = 0for i in data_list:img = cv2.imread(data_path+'/'+i, cv2.IMREAD_GRAYSCALE)#读取灰度图片img = cv2.resize(img, (ROWS, COLS), interpolation=cv2.INTER_CUBIC)img = np.array(img)data_label.append(i[0])#标签列表data[n] = img            n = n+1labels = np_utils.to_categorical(data_label,10)#转化成onehot编码return data,labels,data_listpickle_file = 'Number2.pickle'#若是想重新清洗数据，则更改pickle_file的内容if not os.path.isfile(pickle_file):  #不存在已经处理过的数据train_data,train_label,train_name = process_img(train_path)test_data,test_label,test_name = process_img(test_path)print('Saving data to pickle file...')try:with open(pickle_file, 'wb') as pfile:pickle.dump({'train_data': train_data,'test_data': test_data,'train_label': train_label,'test_label': test_label,'train_name': train_name,'test_name': test_name,},pfile, pickle.HIGHEST_PROTOCOL)except Exception as e:print('Unable to save data to', pickle_file, ':', e)raiseprint('Data cached in pickle file.')
else:#存在文件，直接读取# 读取数据with open(pickle_file, 'rb') as f:pickle_data = pickle.load(f)       # 反序列化，与pickle.dump相反train_data = pickle_data['train_data']train_label = pickle_data['train_label']train_name = pickle_data['train_name']test_data = pickle_data['test_data']test_label = pickle_data['test_label']test_name = pickle_data['test_name']del pickle_data  # 释放内存print('Data and modules loaded.')train_data = train_data.reshape((8000, 28,28,1)).astype('float32') / 255
test_data = test_data.reshape((1200, 28,28,1)).astype('float32') / 255
print(train_data.shape)
print(test_data.shape)

模型搭建与运行

def run(train_data,train_label):batch_size = 256max_epochs = 45model = Sequential()#增加一个卷积层，卷积核数为64，大小为3*3，步长为1model.add(Conv2D(filters=64,kernel_size=(3,3),strides=(1,1),padding = 'same',input_shape=(28,28,1),activation='relu'))#增加一个最大池化层,大小为2*2model.add(MaxPooling2D(pool_size=(2,2)))使用了三个卷积层，两个Dropoutmodel.add(Conv2D(filters=128,kernel_size=(3,3),strides=(1,1),padding='same',activation='relu'))model.add(MaxPooling2D(pool_size=(2,2)))model.add(Dropout(0.5))model.add(Conv2D(filters=256,kernel_size=(3,3),strides=(1,1),padding='same',activation='relu'))    model.add(MaxPooling2D(pool_size=(2,2)))model.add(Dropout(0.5))model.add(Flatten())#把当前层节点展平model.add(Dense(256,activation='relu'))model.add(Dense(10,activation='softmax'))model.compile(loss = 'categorical_crossentropy',optimizer='adagrad',metrics=['accuracy'])#'adagrad'  Adam,SGD,RMSprop#最后定义损失函数early_stopping = EarlyStopping(monitor='val_loss', patience=3, verbose=1, mode='auto')model.fit(train_data,train_label,batch_size=batch_size,epochs=max_epochs,validation_split=0.2,verbose=2,shuffle=True,callbacks=[early_stopping])return modelmymodel = run(train_data,train_label)

预测

scores = mymodel.evaluate(test_data,test_label,verbose=0)
print('\nTest Acc:%.2f%%' % (scores[1] * 100))predict_label = mymodel.predict(test_data,verbose = 0)
classes = []
for i in range(len(predict_label)):  #转换成  max_value=max(predict_label[i])        for j in range(len(predict_label[i])):         if max_value == predict_label[i][j]:               predict_label[i][j] = 1 classes.append(j)else:                predict_label[i][j] = 0
cor = 0
for i in range(len(predict_label)): print(test_name[i],classes[i])if(int(test_name[i][0]) == classes[i]):cor = cor+1
print(cor,cor/len(predict_label))

结果

准确率 0.9775左右

完整代码

import os,cv2,random
import zipfile
import numpy as np
from PIL import Image,ImageQt,ImageDraw,ImageFont
import matplotlib
import matplotlib.pyplot as plt
from keras.utils import np_utils
import picklefrom keras.layers import Input,Dense,Dense,Conv2D,Dropout,MaxPooling2D,Flatten,Activation,AveragePooling2D
from keras.optimizers import Adam,SGD,RMSprop,Adadelta
from keras.callbacks import Callback,EarlyStopping,ReduceLROnPlateau
from keras.preprocessing.image import ImageDataGenerator
from keras.applications.vgg16 import VGG16
from keras.applications.resnet50 import ResNet50
from keras.models import Model,Sequential,load_model
from keras.utils import np_utilsfrom sklearn.metrics import accuracy_scoretrain_name = 'train-8000-clean.zip'
test_name = 'test1--1200-clean.zip'
train = zipfile.ZipFile(train_name)
test= zipfile.ZipFile(test_name)train_path = './tain_data'
test_path = './test_data'if not os.path.exists(train_path):os.makedirs(train_path)for name in train.namelist():train.extract(name,train_path)
train.close() 
if not os.path.exists(test_path):os.makedirs(test_path)for name in test.namelist():test.extract(name,test_path)
test.close()ROWS = 28
COLS = 28
CHANNELS = 1
def process_img(data_path):data_list = []data_label = []for i in os.listdir(data_path):data_list.append(i)count = len(data_list)random.shuffle(data_list)#打乱图片名称data = np.ndarray((count, ROWS, COLS), dtype=np.uint8)n = 0for i in data_list:img = cv2.imread(data_path+'/'+i, cv2.IMREAD_GRAYSCALE)#读取灰度图片img = cv2.resize(img, (ROWS, COLS), interpolation=cv2.INTER_CUBIC)img = np.array(img)data_label.append(i[0])#标签列表data[n] = img            n = n+1labels = np_utils.to_categorical(data_label,10)#转化成onehot编码return data,labels,data_listpickle_file = 'Number2.pickle'#若是想重新清洗数据，则更改pickle_file的内容if not os.path.isfile(pickle_file):  #不存在已经处理过的数据train_data,train_label,train_name = process_img(train_path)test_data,test_label,test_name = process_img(test_path)print('Saving data to pickle file...')try:with open(pickle_file, 'wb') as pfile:pickle.dump({'train_data': train_data,'test_data': test_data,'train_label': train_label,'test_label': test_label,'train_name': train_name,'test_name': test_name,},pfile, pickle.HIGHEST_PROTOCOL)except Exception as e:print('Unable to save data to', pickle_file, ':', e)raiseprint('Data cached in pickle file.')
else:#存在文件，直接读取# 读取数据with open(pickle_file, 'rb') as f:pickle_data = pickle.load(f)       # 反序列化，与pickle.dump相反train_data = pickle_data['train_data']train_label = pickle_data['train_label']train_name = pickle_data['train_name']test_data = pickle_data['test_data']test_label = pickle_data['test_label']test_name = pickle_data['test_name']del pickle_data  # 释放内存print('Data and modules loaded.')train_data = train_data.reshape((8000, 28,28,1)).astype('float32') / 255
test_data = test_data.reshape((1200, 28,28,1)).astype('float32') / 255
print(train_data.shape)
print(test_data.shape)def run(train_data,train_label):batch_size = 256max_epochs = 45model = Sequential()#增加一个卷积层，卷积核数为64，大小为3*3，步长为1model.add(Conv2D(filters=64,kernel_size=(3,3),strides=(1,1),padding = 'same',input_shape=(28,28,1),activation='relu'))#增加一个最大池化层,大小为2*2model.add(MaxPooling2D(pool_size=(2,2)))model.add(Conv2D(filters=128,kernel_size=(3,3),strides=(1,1),padding='same',activation='relu'))model.add(MaxPooling2D(pool_size=(2,2)))model.add(Dropout(0.5))model.add(Conv2D(filters=256,kernel_size=(3,3),strides=(1,1),padding='same',activation='relu'))    model.add(MaxPooling2D(pool_size=(2,2)))model.add(Dropout(0.5))model.add(Flatten())#把当前层节点展平model.add(Dense(256,activation='relu'))model.add(Dense(10,activation='softmax'))model.compile(loss = 'categorical_crossentropy',optimizer='adagrad',metrics=['accuracy'])#'adagrad'  Adam,SGD,RMSprop#最后定义损失函数early_stopping = EarlyStopping(monitor='val_loss', patience=3, verbose=1, mode='auto')model.fit(train_data,train_label,batch_size=batch_size,epochs=max_epochs,validation_split=0.2,verbose=2,shuffle=True,callbacks=[early_stopping])return modelmymodel = run(train_data,train_label)#最后放入批量样本，进行训练
scores = mymodel.evaluate(test_data,test_label,verbose=0)
print('\nTest Acc:%.2f%%' % (scores[1] * 100))
#查看预测的每一条结果，如果有强化学习的需要可以再次学习
predict_label = mymodel.predict(test_data,verbose = 0)
classes = []
for i in range(len(predict_label)):max_value=max(predict_label[i])        for j in range(len(predict_label[i])):         if max_value == predict_label[i][j]:               predict_label[i][j] = 1 classes.append(j)else:                predict_label[i][j] = 0
cor = 0
for i in range(len(predict_label)): print(test_name[i],classes[i])if(int(test_name[i][0]) == classes[i]):cor = cor+1
print(cor,cor/len(predict_label))