ZYNQ图像处理项目——线性神经网络识别mnist

一、线性神经网络识别mnist

线性神经网络其实也可以叫做线性分类器，其实就是没有激活函数的神经网络来对目标进行识别，像支持向量机、多元回归等都是线性的。这边我采用了线性神经网络来识别mnist数字。
我这边是看了一本讲神经网络的书籍，然后打算从线性神经网络入手，实现手写数字的识别。这本书真的讲的很通俗易懂，大家可以去看看，书名如下：

二、matlab训练mnist数据集

matlab的主要代码如下图所示，实现了mnist数据的读取、训练、测试和权重保存。这边不展开讲代码，matlab相关的代码我上传到资源了，大家如果想要看的就可以直接下载。

https://download.csdn.net/download/qq_40995480/87098176

clc;
% clear all;

Images = loadMNISTImages('./MNIST/t10k-images.idx3-ubyte');
Images = reshape(Images, 28, 28, []);
for i=1:10000
    Images(:,:,i) = Images(:,:,i)';
end
Images = reshape(Images, 784, 1, []);
Labels = loadMNISTLabels('./MNIST/t10k-labels.idx1-ubyte');
Labels(Labels == 0) = 10;    % 0 --> 10

rng(1);

W=1e-2*(2*rand(10,784)-1);

X = Images(:, :, 1:8000);
D = Labels(1:8000);

for epoch = 1:50
  epoch;
  [W] = MnistSoftmax(W, X, D);
end

% Test
X = Images(:, :, 8001:10000);
D = Labels(8001:10000);

acc = 0;
N   = length(D);
for k = 1:N
  x = X(:, :, k);                   % Input,           784x1
  v=W*x;
  y=softmax(v);

  [~, i] = max(y);
  if i == D(k)
    acc = acc + 1;
  end
end

%real test
number = imread('./MNIST/number9.bmp');
number=double(number)/15;
number=reshape(number,28,28);
number=number';
number=reshape(number,784,1);

v=W*number;
y=Softmax(v);

acc = acc / N;
fprintf('Accuracy is %f\n', acc);

save the W
W=W/max(max(W))*8191;
W=round(W);
save('weight.mat','W');

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

三、matlab实现线性神经网络

用训练好的权重系数进行书写数字识别，下面是matlab实现的代码，主要干的事情就是对图像的预处理、提取28*28的图像矩阵，之后利用权重与数据乘积来判断最大可能的数字。这部分代码大家也可以到我的资源下载。

https://download.csdn.net/download/qq_40995480/87098187

tic
clc;
clear all;

%read picture
%height and width
img_rgb=imread('./pic/number0.bmp');
h=size(img_rgb,1);
w=size(img_rgb,2);
%get rgb picture
figure(1);
subplot(121);
imshow(img_rgb);
title('rgb picture');
%get gray picture
% Relized method:myself Algorithm realized
% Y = ( R*77 + G*150 + B*29) >>8
% Cb = (-R*44 - G*84 + B*128) >>8
% Cr = ( R*128 - G*108 - B*20) >>8
img_rgb=double(img_rgb);
img_y=zeros(h,w);
img_u=zeros(h,w);
img_v=zeros(h,w);
for i = 1 : h
    for j = 1 : w
        img_y(i,j) = bitshift(( img_rgb(i,j,1)*77 + img_rgb(i,j,2)*150 + img_rgb(i,j,3)*29),-8);
        img_u(i,j) = bitshift((-img_rgb(i,j,1)*44 - img_rgb(i,j,2)*84 + img_rgb(i,j,3)*128 + 32678),-8);
        img_v(i,j) = bitshift(( img_rgb(i,j,1)*128 - img_rgb(i,j,2)*108 - img_rgb(i,j,3)*20 + 32678),-8);
    end
end
img_y = uint8(img_y); 
img_u = uint8(img_u); 
img_v = uint8(img_v); 
img_rgb = uint8(img_rgb);
subplot(122);
imshow(img_y);
title('gray picture');

%get bin picture
img_y=double(img_y);
img_thresh=zeros(h,w);
img_thresh_inverse=zeros(h,w);
THRESH_HOLD=150;
for i=1:h
   for j=1:w
       if(i>=384-56 && i<=384+55 && j>=512-56 && j<=512+55)
           if(img_y(i,j)>THRESH_HOLD)
               img_thresh(i,j)=255;
               img_thresh_inverse(i,j)=0;
           else 
               img_thresh(i,j)=0;
               img_thresh_inverse(i,j)=255;
           end
       else 
           img_thresh(i,j)=img_y(i,j);
           img_thresh_inverse(i,j)=img_y(i,j);
       end
   end
end
img_y=uint8(img_y);
img_thresh=uint8(img_thresh);
img_thresh_inverse=uint8(img_thresh_inverse);
figure(2);
subplot(121);
imshow(img_thresh);
title('bin picture');
subplot(122);
imshow(img_thresh_inverse);
title('bin inverse picture');

%open operation:corrosion and expansion
img_corrosion=zeros(h,w);
for i=1:h
   for j=1:w
       if(i>=384-56 && i<=384+55 && j>=512-56 && j<=512+55)
           a=img_thresh_inverse(i-1,j-1)&img_thresh_inverse(i-1,j)&img_thresh_inverse(i-1,j+1);
           b=img_thresh_inverse(i,j-1)&img_thresh_inverse(i,j)&img_thresh_inverse(i,j+1);
           c=img_thresh_inverse(i+1,j-1)&img_thresh_inverse(i+1,j)&img_thresh_inverse(i+1,j+1);
           if((a&b&c)==1)
               img_corrosion(i,j)=255;
           else 
               img_corrosion(i,j)=0;
           end
       else
               img_corrosion(i,j)=img_thresh_inverse(i,j);
       end
   end
end
img_corrosion=uint8(img_corrosion);
figure(3);
subplot(121);
imshow(img_corrosion);
title('corrosion picture');
img_expansion=zeros(h,w);
for i=1:h
   for j=1:w
       if(i>=384-55 && i<=384+54 && j>=512-55 && j<=512+54)
           a=img_corrosion(i-1,j-1)|img_corrosion(i-1,j)|img_corrosion(i-1,j+1);
           b=img_corrosion(i,j-1)|img_corrosion(i,j)|img_corrosion(i,j+1);
           c=img_corrosion(i+1,j-1)|img_corrosion(i+1,j)|img_corrosion(i+1,j+1);
           if((a|b|c)==1)
               img_expansion(i,j)=255;
           else 
               img_expansion(i,j)=0;
           end
       else
               img_expansion(i,j)=img_corrosion(i,j);
       end
   end
end
img_expansion=uint8(img_expansion);
subplot(122);
imshow(img_expansion);
title('expansion picture');

%get number matrix
test=imread('./pic/output_file1.bmp');
number=zeros(28,28);
for i=1:28
    for j=1:28
        number(i,j)=sum(sum(test(384-56+4*(i-1):384-56+4*i-1 , 512-56+4*(j-1):512-56+4*j-1)))/16;
    end
end
number=floor(number);
number=uint8(number);
pic=load('./pic/pic_28x28.txt');
pic=reshape(pic,28,28);
pic=pic';
pic=uint8(pic);
figure(4);
subplot(121);
imshow(number)
title('matlab number');
subplot(122);
imshow(pic)
title('vivado number');


%use softmax to recog the number
load('weight');
number=double(number);
number=number';
number=reshape(number,784,1);
v=W*number;
v=v/max(v)*10;
y=softmax(v);
n=0;
for i=1:10
   if(y(i)>=n)
       n=y(i);
       number_recog=i;
   end
   if(number_recog==10)
      number_recog=0; 
   end
end
fprintf('the probability is %f\n',n*100);
fprintf('the number is %d\n',number_recog);
toc

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
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160

下面两张图是我测试的原始图像和转灰度之后的图像。

下图是将灰度图像对想要处理的区域(数字区域)进行二值化，之后在翻转。

下面这两张图分别是我用matlab和vivado仿真提取得到的28*28的数据。

四、ZYNQ实现线性神经网络

关于ZYNQ实现神经网络的思路其实和matlab写的代码的思路是一样的，就是将其用verilog来实现而已，下图是我将线性神经网络部署到ZYNQ上实现的具体效果。在测试中发现，线性神经网络其实性能相较于卷积神经网络要差一点，下面工作会将激活函数、卷积层、池化层这些内容加进去，用ZYNQ来实现卷积神经网络。