简 介： 对于七段数码管图片进行增强，利用LeNet对数据集建立识别模型。基于这个模型将来用于实际图片中的数字进行识别。

关键词： LCD，LENET

§01 七段数码识别

1.1 数据集合准备

1.1.1 论文中图片

（1）原始图片

  如下是来自于 7 segment LCD Optical Character Recognition 的十个字符的数据集合。将它们分割成各自的图片。

（2）分割数字图片

from headm import *                 # =
import cv2

iddim = [8, 11, 9, 10]

outdir = r'd:\temp\lcd2seg'

if not os.path.isdir(outdir):
    os.makedirs(outdir)

def pic2num(picid, boxid, rownum, headstr):
    global outdir

    imgfile = tspgetdopfile(picid)
    printt(imgfile:)

    img = cv2.imread(imgfile)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    printt(img.shape, gray.shape)

    plt.imshow(gray, plt.cm.gray)
    plt.savefig(r'd:\temp\figure1.jpg')
    tspshowimage(image=r'd:\temp\figure1.jpg')

    imgrect = tspgetrange(picid)
    boxrect = tspgetrange(boxid)

    imgwidth = gray.shape[1]
    imgheight = gray.shape[0]

    boxheight = boxrect[3] - boxrect[1]
    boxwidth = boxrect[2] - boxrect[0]
    for i in range(rownum):
        rowstart = boxrect[1] - imgrect[1] + boxheight * i / rownum
        rowend = boxrect[1] - imgrect[1] + boxheight * (i+1) / rownum

        imgtop = int(rowstart * imgheight / (imgrect[3] - imgrect[1]))
        imgbottom = int(rowend * imgheight / (imgrect[3] - imgrect[1]))

        for j in range(10):
            colstart = boxrect[0] - imgrect[0] + boxwidth * j / 10
            colend = boxrect[0] - imgrect[0] + boxwidth * (j + 1) / 10

            imgleft = int(colstart * imgwidth / (imgrect[2] - imgrect[0]))
            imgright = int(colend * imgwidth / (imgrect[2] - imgrect[0]))

            imgdata = gray[imgtop:imgbottom+1, imgleft:imgright+1]

            fn = os.path.join(outdir, '%s_%d_%d.jpg'%(headstr, i, j))
            cv2.imwrite(fn, imgdata)

pic2num(iddim[2], iddim[3], 4, '2')

  

 

  
1
  
2
  
3
  
4
  
5
  
6
  
7
  
8
  
9
  
10
  
11
  
12
  
13
  
14
  
15
  
16
  
17
  
18
  
19
  
20
  
21
  
22
  
23
  
24
  
25
  
26
  
27
  
28
  
29
  
30
  
31
  
32
  
33
  
34
  
35
  
36
  
37
  
38
  
39
  
40
  
41
  
42
  
43
  
44
  
45
  
46
  
47
  
48
  
49
  
50
  
51
  
52
 

  总共获得图片70个。图片的命名标准：最后一个字母代表数字。

1.2 数据集合扩增

  对图片进行扩增，扩增方法：

• 旋转：-25 ~ 25
• 比例：0.9 ~ 1.1
• 平移：x,y: -5,5

  每个图片扩增到原来的225个。

1.2.1 图片扩增程序

from headm import *                 # =
import cv2
import paddle
from paddle.vision.transforms import Resize,rotate,adjust_brightness,adjust_contrast
from PIL import Image
from tqdm import tqdm
from tqdm import tqdm

indir = '/home/aistudio/work/lcd7seg/lcd2seg'
outdir = '/home/aistudio/work/lcd7seg/lcdaugment'

infile = os.listdir(indir)
printt(infile:)

BUF_SIZE = 512
def rotateimg(img, degree, ratio, shift_x, shift_y):

    imgavg = mean(img.flatten())
    if imgavg < 100:
        img[img >= 200] = imgavg
        imgavg = mean(img.flatten())
        img[img >= 200] = imgavg
    else: imgavg = 255

    whiteimg = ones((BUF_SIZE, BUF_SIZE, 3))*(imgavg/255)
    imgshape = img.shape
    imgwidth = imgshape[0]
    imgheight = imgshape[1]
    left = (BUF_SIZE - imgwidth) // 2
    top = (BUF_SIZE - imgheight) // 2
    whiteimg[left:left+imgwidth, top:top+imgheight, :] = img/255

    img1 = rotate(whiteimg, degree)

    centerx = BUF_SIZE//2+shift_x
    centery = BUF_SIZE//2+shift_y

    imgwidth1 = int(imgwidth * ratio)
    imgheight1 = int(imgheight * ratio)

    imgleft = centerx - imgwidth1//2
    imgtop = centery - imgheight1//2

    imgnew = img1[imgleft:imgwidth1+imgleft, imgtop:imgtop+imgheight1, :]
    imgnew = cv2.resize(imgnew, (32,32))

    return imgnew*255

gifpath = '/home/aistudio/GIF'
if not os.path.isdir(gifpath):
    os.makedirs(gifpath)
gifdim = os.listdir(gifpath)
for f in gifdim:
    fn = os.path.join(gifpath, f)
    if os.path.isfile(fn):
        os.remove(fn)

count = 0
def outimg(fn):
    global count

    imgfile = os.path.join(indir, fn)
    img = cv2.imread(imgfile)

    num = fn.split('.')[0][-1]

    rotate = linspace(-25, 25, 5).astype(int)
    ratio = linspace(0.8, 1.2, 5)
    shiftxy = linspace(-5, 5, 3).astype(int)

    for r in rotate:
        for t in ratio:
            for s in shiftxy:
                for h in shiftxy:
                    imgout = rotateimg(img, r, t, s, h)
                    outfilename = os.path.join(gifpath, '%04d_%s.jpg'%(count, num))
                    count += 1
                    cv2.imwrite(outfilename, imgout)
                    outfilename = os.path.join(gifpath, '%04d_%s.jpg'%(count, num))
                    count += 1
                    cv2.imwrite(outfilename, 255-imgout)

for id,f in tqdm(enumerate(infile)):
    if f.find('jpg') < 0: continue
    outimg(f)

  

 

  
1
  
2
  
3
  
4
  
5
  
6
  
7
  
8
  
9
  
10
  
11
  
12
  
13
  
14
  
15
  
16
  
17
  
18
  
19
  
20
  
21
  
22
  
23
  
24
  
25
  
26
  
27
  
28
  
29
  
30
  
31
  
32
  
33
  
34
  
35
  
36
  
37
  
38
  
39
  
40
  
41
  
42
  
43
  
44
  
45
  
46
  
47
  
48
  
49
  
50
  
51
  
52
  
53
  
54
  
55
  
56
  
57
  
58
  
59
  
60
  
61
  
62
  
63
  
64
  
65
  
66
  
67
  
68
  
69
  
70
  
71
  
72
  
73
  
74
  
75
  
76
  
77
  
78
  
79
  
80
  
81
  
82
  
83
  
84
  
85
 

1.2.2 扩增结果

  每个图片扩增后为225，加上颜色反向，70个图片最终获得数字图片为 31500个图片。

数据库参数：

数字： 0 ~ 9
个数： 31500
尺寸： 32×32×3

1.3 存储数据文件

1.3.1 转换代码

from headm import *                 # =
import cv2

lcddir = '/home/aistudio/work/lcd7seg/GIF'
filedim = os.listdir(lcddir)

printt(len(filedim))
imagedim = []
labeldim = []

for f in filedim:
    fn = os.path.join(lcddir, f)
    if fn.find('.jpg') < 0: continue

    gray = cv2.cvtColor(cv2.imread(fn), cv2.COLOR_BGR2GRAY)

    imagedim.append(gray)
    num = int(f.split('.')[0][-1])
    labeldim.append(num)

imgarray = array(imagedim)
label = array(labeldim)

savez('/home/aistudio/work/lcddata', lcd=imgarray, label=label)

printt("Save lcd data.")

  

 

  
1
  
2
  
3
  
4
  
5
  
6
  
7
  
8
  
9
  
10
  
11
  
12
  
13
  
14
  
15
  
16
  
17
  
18
  
19
  
20
  
21
  
22
  
23
  
24
  
25
  
26
 

1.3.2 数据文件

数据文件：

文件名称：lcddata.npz
lcd： (31500,32,32)
label：(31500)

§02 构建LeNet

2.1 构建LeNet网络

from headm import *                 # =

import paddle
import paddle.fluid as fluid
from paddle import to_tensor as TT
from paddle.nn.functional import square_error_cost as SQRC

datafile = '/home/aistudio/work/lcddata.npz'
data = load(datafile)
lcd = data['lcd']
llabel = data['label']
printt(lcd.shape, llabel.shape)

class Dataset(paddle.io.Dataset):
    def __init__(self, num_samples):
        super(Dataset, self).__init__()
        self.num_samples = num_samples

    def __getitem__(self, index):
        data = lcd[index][newaxis,:,:]
        label = llabel[index]
        return paddle.to_tensor(data,dtype='float32'), paddle.to_tensor(label,dtype='int64')

    def __len__(self):
        return self.num_samples

_dataset = Dataset(len(llabel))
train_loader = paddle.io.DataLoader(_dataset, batch_size=100, shuffle=True)

imgwidth = 32
imgheight = 32
inputchannel = 1
kernelsize   = 5
targetsize = 10
ftwidth = ((imgwidth-kernelsize+1)//2-kernelsize+1)//2
ftheight = ((imgheight-kernelsize+1)//2-kernelsize+1)//2

class lenet(paddle.nn.Layer):
    def __init__(self, ):
        super(lenet, self).__init__()
        self.conv1 = paddle.nn.Conv2D(in_channels=inputchannel, out_channels=6, kernel_size=kernelsize, stride=1, padding=0)
        self.conv2 = paddle.nn.Conv2D(in_channels=6, out_channels=16, kernel_size=kernelsize, stride=1, padding=0)
        self.mp1    = paddle.nn.MaxPool2D(kernel_size=2, stride=2)
        self.mp2    = paddle.nn.MaxPool2D(kernel_size=2, stride=2)
        self.L1     = paddle.nn.Linear(in_features=ftwidth*ftheight*16, out_features=120)
        self.L2     = paddle.nn.Linear(in_features=120, out_features=86)
        self.L3     = paddle.nn.Linear(in_features=86, out_features=targetsize)

    def forward(self, x):
        x = self.conv1(x)
        x = paddle.nn.functional.relu(x)
        x = self.mp1(x)
        x = self.conv2(x)
        x = paddle.nn.functional.relu(x)
        x = self.mp2(x)
        x = paddle.flatten(x, start_axis=1, stop_axis=-1)
        x = self.L1(x)
        x = paddle.nn.functional.relu(x)
        x = self.L2(x)
        x = paddle.nn.functional.relu(x)
        x = self.L3(x)
        return x

model = lenet()

optimizer = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())
def train(model):
    model.train()
    epochs = 100
    for epoch in range(epochs):
        for batch, data in enumerate(train_loader()):
            out = model(data[0])
            loss = paddle.nn.functional.cross_entropy(out, data[1])
            acc = paddle.metric.accuracy(out, data[1])
            loss.backward()
            optimizer.step()
            optimizer.clear_grad()

        print('Epoch:{}, Accuracys:{}'.format(epoch, acc.numpy()))

train(model)

paddle.save(model.state_dict(), './work/model.pdparams')

  

 

  
1
  
2
  
3
  
4
  
5
  
6
  
7
  
8
  
9
  
10
  
11
  
12
  
13
  
14
  
15
  
16
  
17
  
18
  
19
  
20
  
21
  
22
  
23
  
24
  
25
  
26
  
27
  
28
  
29
  
30
  
31
  
32
  
33
  
34
  
35
  
36
  
37
  
38
  
39
  
40
  
41
  
42
  
43
  
44
  
45
  
46
  
47
  
48
  
49
  
50
  
51
  
52
  
53
  
54
  
55
  
56
  
57
  
58
  
59
  
60
  
61
  
62
  
63
  
64
  
65
  
66
  
67
  
68
  
69
  
70
  
71
  
72
  
73
  
74
  
75
  
76
  
77
  
78
  
79
  
80
  
81
  
82
  
83
 

2.2 训练结果

2.2.1 第一次训练

训练参数：

Lr=0.001：
BatchSize=100：

2.2.2 第二次训练

训练参数：

BatchSize=1000：
Lr=0.001：

2.2.3 第三次训练

训练参数：

BatchSize=5000：
Lr=0.005：

2.2.4 第四次训练

训练参数：

BatchSize：2000
Lr=0.002：

2.2.5 第五次训练

训练参数：

BatchSize：2000
Lr=0.001：

※ 处理总结 ※

  对于七段数码管图片进行增强，利用LeNet对数据集建立识别模型。基于这个模型将来用于实际图片中的数字进行识别。

后期实验

对于通过少量的样本进行变形之后所获得的模型，在 测试LCDNet对于万用表读数识别效果 中进行了测试，发现效果好很不理想，主要原因：

• 训练样本太少，仅仅通过原有的模板进行初步图片变形，还是无法适应大多数的情况；这说明图片的扩增与实际的分布之间还是有很大的差别。

测试实际图片

• 对于样本的归一化，Normalization，非常重要。在 一个中等规模的七段数码数据库以及利用它训练的识别网络 引入了归一化，使得训练输出的模型适应度得到了很大的提高。

3.1 附件：处理程序

3.1.1 LCDLENET

from headm import *                 # =

import paddle
import paddle.fluid as fluid
from paddle import to_tensor as TT
from paddle.nn.functional import square_error_cost as SQRC

datafile = '/home/aistudio/work/lcddata.npz'
data = load(datafile)
lcd = data['lcd']
llabel = data['label']
printt(lcd.shape, llabel.shape)

class Dataset(paddle.io.Dataset):
    def __init__(self, num_samples):
        super(Dataset, self).__init__()
        self.num_samples = num_samples

    def __getitem__(self, index):
        data = lcd[index][newaxis,:,:]
        label = llabel[index]
        return paddle.to_tensor(data,dtype='float32'), paddle.to_tensor(label,dtype='int64')

    def __len__(self):
        return self.num_samples

_dataset = Dataset(len(llabel))
train_loader = paddle.io.DataLoader(_dataset, batch_size=2000, shuffle=True)

imgwidth = 32
imgheight = 32
inputchannel = 1
kernelsize   = 5
targetsize = 10
ftwidth = ((imgwidth-kernelsize+1)//2-kernelsize+1)//2
ftheight = ((imgheight-kernelsize+1)//2-kernelsize+1)//2

class lenet(paddle.nn.Layer):
    def __init__(self, ):
        super(lenet, self).__init__()
        self.conv1 = paddle.nn.Conv2D(in_channels=inputchannel, out_channels=6, kernel_size=kernelsize, stride=1, padding=0)
        self.conv2 = paddle.nn.Conv2D(in_channels=6, out_channels=16, kernel_size=kernelsize, stride=1, padding=0)
        self.mp1    = paddle.nn.MaxPool2D(kernel_size=2, stride=2)
        self.mp2    = paddle.nn.MaxPool2D(kernel_size=2, stride=2)
        self.L1     = paddle.nn.Linear(in_features=ftwidth*ftheight*16, out_features=120)
        self.L2     = paddle.nn.Linear(in_features=120, out_features=86)
        self.L3     = paddle.nn.Linear(in_features=86, out_features=targetsize)

    def forward(self, x):
        x = self.conv1(x)
        x = paddle.nn.functional.relu(x)
        x = self.mp1(x)
        x = self.conv2(x)
        x = paddle.nn.functional.relu(x)
        x = self.mp2(x)
        x = paddle.flatten(x, start_axis=1, stop_axis=-1)
        x = self.L1(x)
        x = paddle.nn.functional.relu(x)
        x = self.L2(x)
        x = paddle.nn.functional.relu(x)
        x = self.L3(x)
        return x

model = lenet()

optimizer = paddle.optimizer.Adam(learning_rate=0.001, parameters=model.parameters())
def train(model):
    model.train()
    epochs = 200
    for epoch in range(epochs):
        for batch, data in enumerate(train_loader()):
            out = model(data[0])
            loss = paddle.nn.functional.cross_entropy(out, data[1])
            acc = paddle.metric.accuracy(out, data[1])
            loss.backward()
            optimizer.step()
            optimizer.clear_grad()

        printt('Epoch:{}, Accuracys:{}'.format(epoch, acc.numpy()))

train(model)

paddle.save(model.state_dict(), './work/model3.pdparams')

filename = '/home/aistudio/stdout.txt'

accdim = []
with open(filename, 'r') as f:
    for l in f.readlines():
        ll = l.split('[')
        if len(ll) < 2: continue
        ll = ll[-1].split(']')
        if len(ll) < 2: continue

        accdim.append(float(ll[0]))

plt.figure(figsize=(12, 8))
plt.plot(accdim)
plt.xlabel("Step")
plt.ylabel("Acc")
plt.grid(True)
plt.tight_layout()
plt.show()

  

 

  
1
  
2
  
3
  
4
  
5
  
6
  
7
  
8
  
9
  
10
  
11
  
12
  
13
  
14
  
15
  
16
  
17
  
18
  
19
  
20
  
21
  
22
  
23
  
24
  
25
  
26
  
27
  
28
  
29
  
30
  
31
  
32
  
33
  
34
  
35
  
36
  
37
  
38
  
39
  
40
  
41
  
42
  
43
  
44
  
45
  
46
  
47
  
48
  
49
  
50
  
51
  
52
  
53
  
54
  
55
  
56
  
57
  
58
  
59
  
60
  
61
  
62
  
63
  
64
  
65
  
66
  
67
  
68
  
69
  
70
  
71
  
72
  
73
  
74
  
75
  
76
  
77
  
78
  
79
  
80
  
81
  
82
  
83
  
84
  
85
  
86
  
87
  
88
  
89
  
90
  
91
  
92
  
93
  
94
  
95
  
96
  
97
  
98
  
99
  
100
  
101
  
102
  
103
 

3.1.2 DrawCurve

accdim = []
with open(filename, 'r') as f:
    for l in f.readlines():
        ll = l.split('[')
        if len(ll) < 2: continue
        ll = ll[-1].split(']')
        if len(ll) < 2: continue

        accdim.append(float(ll[0]))

plt.figure(figsize=(12, 8))
plt.plot(accdim)
plt.xlabel("Step")
plt.ylabel("Acc")
plt.grid(True)
plt.tight_layout()
plt.show()

  

 

  
1
  
2
  
3
  
4
  
5
  
6
  
7
  
8
  
9
  
10
  
11
  
12
  
13
  
14
  
15
  
16
  
17
 

3.1.3 数据处理程序

from headm import *                 # =
import cv2

lcddir = '/home/aistudio/work/lcd7seg/GIF'
filedim = os.listdir(lcddir)

printt(len(filedim))
imagedim = []
labeldim = []

for f in filedim:
    fn = os.path.join(lcddir, f)
    if fn.find('.jpg') < 0: continue

    gray = cv2.cvtColor(cv2.imread(fn), cv2.COLOR_BGR2GRAY)

    imagedim.append(gray)
    num = int(f.split('.')[0][-1])
    labeldim.append(num)

imgarray = array(imagedim)
label = array(labeldim)

savez('/home/aistudio/work/lcddata', lcd=imgarray, label=label)

printt("Save lcd data.")

  

 

  
1
  
2
  
3
  
4
  
5
  
6
  
7
  
8
  
9
  
10
  
11
  
12
  
13
  
14
  
15
  
16
  
17
  
18
  
19
  
20
  
21
  
22
  
23
  
24
  
25
  
26
 

■ 相关文献链接:

• 7 segment LCD Optical Character Recognition

● 相关图表链接:

• 图1.1 七段数码变形图片
• 图1.1.2 分割出的数字图片
• 图 每个图片倍增后的图片
• 图 每个图片倍增后的图片
• 图2.2.1 训练精度
• 图2.2.2 训练过程中的识别精度
• 图2.2.3 训练过程中的识别精度
• 图2.2.4 训练过程中的识别精度
• 图2.2.5 训练过程中的识别精度

文章来源: zhuoqing.blog.csdn.net，作者：卓晴，版权归原作者所有，如需转载，请联系作者。

原文链接：zhuoqing.blog.csdn.net/article/details/122347898