[转]Theano下用CNN(卷积神经网络)做车牌中文字符OCR

Theano下用CNN(卷积神经网络)做车牌中文字符OCR

原文地址：http://m.blog.csdn.net/article/details?id=50989742

之前时间一直在看 Michael Nielsen 先生的 Deep Learning 教程。

用了他的代码在theano下测试了下中文车牌字符的识别。由于我没有GPU，简单的在进行了16个epoch之后，识别率达到了 98.41% ，由于图像本来质量就不高，达到这个识别率，效果挺不错了。

一共 31 类车牌中文字符数据来源于中文车牌识别项目 EasyPR 的数据集 . 由于数据集分布很不均匀。可能会导致个别类别拟合不一致，而降低识别率。所以使用随机轻微扭曲图像的方式来生成新的数据以保证数据集各个类目的数量的均衡。

下面是用于轻微扭曲图像来生成更多样本的函数。

def rotRandrom(img,factor,size):

    """ 使图像轻微的畸变

        img 输入图像

        factor 畸变的参数

        size 为图片的目标尺寸

    """

    img = img.reshape(size);

    shape = size;

    pts1 = np.float32([[0,0],[0,shape[0]],[shape[1],0],[shape[1],shape[0]]])

    pts2 = np.float32([[r(factor),r(factor)],[0,shape[0]-r(factor)],[shape[1]-r(factor),0],[shape[1]-r(factor),shape[0]-r(factor)]])

    M  = cv2.getPerspectiveTransform(pts1,pts2);

    dst = cv2.warpPerspective(img,M,(shape[0],shape[1]));

    return dst.ravel();

在训练的时候使用的CNN结构如下

激活函数都为 ReLu

Conv(kernel size 5*5 ) * 25个 feature map->

Pooling 2*2 ->

Conv(kernel size 5*5) * 16个feature map->

Pooling 2*2 ->

FullConnectedLayer 120 个 Neurons ->

FullConnectedLayer 84个 Neurons ->

Softmax Output 31类

import network3

from network3 import Network

from network3 import ConvPoolLayer, FullyConnectedLayer, SoftmaxLayer

from network3 import ReLU

training_data, validation_data, test_data= network3.load_data_cPickle("./data.pkl")

mini_batch_size = 10

net = Network([

        ConvPoolLayer(image_shape=(mini_batch_size, 1, 28, 28),

                      filter_shape=(25, 1, 5, 5),

                      poolsize=(2, 2),

                      activation_fn=ReLU),

        ConvPoolLayer(image_shape=(mini_batch_size, 25, 12, 12),

                      filter_shape=(16, 25, 5, 5),

                      poolsize=(2, 2),

                      activation_fn=ReLU),

        FullyConnectedLayer(n_in=16*4*4, n_out=120, activation_fn=ReLU),

         FullyConnectedLayer(n_in=120, n_out=84, activation_fn=ReLU),

        SoftmaxLayer(n_in=84, n_out=31)], mini_batch_size )

net.SGD(training_data, 60, mini_batch_size, 0.03, validation_data, test_data, lmbda=0.1)

这个函数用于制作自己的数据。

def make_dataset(dirn):

    set = [];

    labels = [] ;

    def findinside(dirname,code,):

        print "code",code;

        print "dirname",dirname;

        for parent,dirnames,filenames in os.walk(dirname):

            adder = 1400 -  len(filenames)

            len_d = len(filenames)

            for filename in filenames:

                path  =parent+"/"+filename

                if(path.endswith(".jpg")):

                            img = cv2.imread(path,cv2.CV_LOAD_IMAGE_GRAYSCALE);

                            img = cv2.resize(img,(28,28));

                            img = img.astype(np.float32)/255;

                            set.append(img.ravel());

                            labels.append(code);

            for i in range(adder):

                c_index = int(np.random.rand() * len_d);

                l_set = len(set)

                set.append(rotrandom.rotRandrom( set[l_set-len_d + c_index],0.88,(28,28)));

                labels.append(code);

        print len(set),dirname,len(filenames)

    for parent,dirnames,filenames in os.walk(dirn):

            num = len(dirnames);

            for i in range(num):

                        c_path = dir_chars + "/"+ dirnames[i];

                        findinside(c_path,i);

    shuffle = np.random.permutation(len(set));

    print len(set)

    set = np.array(set);

    labels = np.array(labels);

    set, labels = set[shuffle], labels[shuffle]

    train_n = int(0.9*len(set))

    training_set,test_set = np.split(set, [train_n])

    training_labels, test_labels = np.split(labels, [train_n])

    print training_labels

    validation_set = test_set.copy();

    validation_labels = test_set.copy();

    training_data = [training_set,training_labels]

    validation_data = [validation_set,validation_labels]

    test_data = [test_set,test_labels]

    data = [ training_data, validation_data, test_data];

    fileid  = open("./data.pkl","wb")

    cPickle.dump(data,fileid)

dir_chars = "./charsChinese"

make_dataset(dir_chars);

在进行了第14个epoch之后获得了 98.41% 训练时间大概在10分钟左右。

Training mini-batch number 0

Training mini-batch number 1000

Training mini-batch number 2000

Training mini-batch number 3000

Epoch 0: validation accuracy 89.15%

This is the best validation accuracy to date.

The corresponding test accuracy is 89.15%

Training mini-batch number 4000

Training mini-batch number 5000

Training mini-batch number 6000

Training mini-batch number 7000

Epoch 1: validation accuracy 94.65%

This is the best validation accuracy to date.

The corresponding test accuracy is 94.65%

Training mini-batch number 8000

Training mini-batch number 9000

Training mini-batch number 10000

Training mini-batch number 11000

Epoch 2: validation accuracy 95.44%

This is the best validation accuracy to date.

The corresponding test accuracy is 95.44%

Training mini-batch number 12000

Training mini-batch number 13000

Training mini-batch number 14000

Training mini-batch number 15000

Epoch 3: validation accuracy 96.13%

This is the best validation accuracy to date.

The corresponding test accuracy is 96.13%

Training mini-batch number 16000

Training mini-batch number 17000

Training mini-batch number 18000

Training mini-batch number 19000

Epoch 4: validation accuracy 96.91%

This is the best validation accuracy to date.

The corresponding test accuracy is 96.91%

Training mini-batch number 20000

Training mini-batch number 21000

Training mini-batch number 22000

Training mini-batch number 23000

Epoch 5: validation accuracy 96.52%

Training mini-batch number 24000

Training mini-batch number 25000

Training mini-batch number 26000

Training mini-batch number 27000

Epoch 6: validation accuracy 96.87%

Training mini-batch number 28000

Training mini-batch number 29000

Training mini-batch number 30000

Training mini-batch number 31000

Epoch 7: validation accuracy 96.87%

Training mini-batch number 32000

Training mini-batch number 33000

Training mini-batch number 34000

Training mini-batch number 35000

Epoch 8: validation accuracy 97.58%

This is the best validation accuracy to date.

The corresponding test accuracy is 97.58%

Training mini-batch number 36000

Training mini-batch number 37000

Training mini-batch number 38000

Training mini-batch number 39000

Epoch 9: validation accuracy 97.49%

Training mini-batch number 40000

Training mini-batch number 41000

Training mini-batch number 42000

Epoch 10: validation accuracy 97.60%

This is the best validation accuracy to date.

The corresponding test accuracy is 97.60%

Training mini-batch number 43000

Training mini-batch number 44000

Training mini-batch number 45000

Training mini-batch number 46000

Epoch 11: validation accuracy 97.93%

This is the best validation accuracy to date.

The corresponding test accuracy is 97.93%

Training mini-batch number 47000

Training mini-batch number 48000

Training mini-batch number 49000

Training mini-batch number 50000

Epoch 12: validation accuracy 97.83%

Training mini-batch number 51000

Training mini-batch number 52000

Training mini-batch number 53000

Training mini-batch number 54000

Epoch 13: validation accuracy 98.04%

This is the best validation accuracy to date.

The corresponding test accuracy is 98.04%

Training mini-batch number 55000

Training mini-batch number 56000

Training mini-batch number 57000

Training mini-batch number 58000

Epoch 14: validation accuracy 98.20%

This is the best validation accuracy to date.

The corresponding test accuracy is 98.20%

Training mini-batch number 59000

Training mini-batch number 60000

Training mini-batch number 61000

Training mini-batch number 62000

Epoch 15: validation accuracy 97.86%

Training mini-batch number 63000

Training mini-batch number 64000

Training mini-batch number 65000

Training mini-batch number 66000

Epoch 16: validation accuracy 98.41%

This is the best validation accuracy to date.

The corresponding test accuracy is 98.41%