[英]OCR with perceptron neural network of Aforge.net answers wrong
I tried to make OCR by perceptrons with Aforge.Net in C#. 我试图用C#中的Aforge.Net用感知器制作OCR。 I learned my network with nine 30*30 pictures in binary. 我用9张30 * 30二进制二进制图片学习了我的网络。 But in the results, it recognizes everything as 'C'. 但是在结果中,它将所有内容识别为“ C”。 this is the code: 这是代码:
private void button1_Click(object sender, EventArgs e)
{
AForge.Neuro.ActivationNetwork network = new AForge.Neuro.ActivationNetwork(new AForge.Neuro.BipolarSigmoidFunction(2), 900, 3);
network.Randomize();
AForge.Neuro.Learning.PerceptronLearning learning = new AForge.Neuro.Learning.PerceptronLearning(network);
learning.LearningRate =1 ;
double[][] input = new double[9][];
for (int i = 0; i < 9; i++)
{
input[i] = new double[900];
}
//Reading A images
for (int i = 1; i <= 3; i++)
{
Bitmap a = AForge.Imaging.Image.FromFile(path + "\\a" + i + ".bmp");
for (int j = 0; j < 30; j++)
for (int k = 0; k < 30; k++)
{
if (a.GetPixel(j, k).ToKnownColor() == KnownColor.White)
{
input[i-1][j * 10 + k] = -1;
}
else
input[i-1][j * 10 + k] = 1;
}
// showImage(a);
}
//Reading B images
for (int i = 1; i <= 3; i++)
{
Bitmap a = AForge.Imaging.Image.FromFile(path + "\\b" + i + ".bmp");
for (int j = 0; j < 30; j++)
for (int k = 0; k < 30; k++)
{
if (a.GetPixel(j , k).ToKnownColor() == KnownColor.White)
{
input[i + 2][j * 10 + k] = -1;
}
else
input[i + 2][j * 10 + k] = 1;
}
// showImage(a);
}
//Reading C images
for (int i = 1; i <= 3; i++)
{
Bitmap a = AForge.Imaging.Image.FromFile(path + "\\c" + i + ".bmp");
for (int j = 0; j < 30; j++)
for (int k = 0; k < 30; k++)
{
if (a.GetPixel(j , k ).ToKnownColor() == KnownColor.White)
{
input[i + 5][j * 10 + k] = -1;
}
else
input[i + 5][j * 10 + k] = 1;
}
// showImage(a);
}
bool needToStop = false;
int iteration = 0;
while (!needToStop)
{
double error = learning.RunEpoch(input, new double[9][] { new double[3] { 1, -1, -1 },new double[3] { 1, -1, -1 },new double[3] { 1, -1, -1 },//A
new double[3] { -1, 1, -1 },new double[3] { -1, 1, -1 },new double[3] { -1, 1, -1 },//B
new double[3] { -1, -1, 1 },new double[3] { -1, -1, 1 },new double[3] { -1, -1, 1 } }//C
/*new double[9][]{ input[0],input[0],input[0],input[1],input[1],input[1],input[2],input[2],input[2]}*/
);
//learning.LearningRate -= learning.LearningRate / 1000;
if (error == 0)
break;
else if (iteration < 1000)
iteration++;
else
needToStop = true;
System.Diagnostics.Debug.WriteLine("{0} {1}", error, iteration);
}
Bitmap b = AForge.Imaging.Image.FromFile(path + "\\b1.bmp");
//Reading A Sample to test Netwok
double[] sample = new double[900];
for (int j = 0; j < 30; j++)
for (int k = 0; k < 30; k++)
{
if (b.GetPixel(j , k ).ToKnownColor() == KnownColor.White)
{
sample[j * 30 + k] = -1;
}
else
sample[j * 30 + k] = 1;
}
foreach (double d in network.Compute(sample))
System.Diagnostics.Debug.WriteLine(d);//Output is Always C = {-1,-1,1}
}
I really wanted to know why it answers wrong. 我真的很想知道为什么它回答错误。
While loading your initial 30x30 images into a double[900] array in the input
structure you are using the following calculation: 在将初始的30x30图像加载到input
结构中的double [900]数组中时,您使用的是以下计算:
for (int j = 0; j < 30; j++)
for (int k = 0; k < 30; k++)
{
if (a.GetPixel(j, k).ToKnownColor() == KnownColor.White)
input[i-1][j * 10 + k] = -1;
else
input[i-1][j * 10 + k] = 1;
}
Your offset calculation is wrong here. 您的偏移量计算在这里是错误的。 You need to change j * 10 + k
to j * 30 + k
or you will get invalid results. 您需要将j * 10 + k
更改为j * 30 + k
否则您将得到无效的结果。 Later you use the correct offset calculation while loading the test image, which is why it's not being matched correctly against the corrupted samples. 稍后,您在加载测试图像时会使用正确的偏移量计算,这就是为什么它无法与损坏的样本正确匹配的原因。
You should write a method to load a bitmap into a double[900]
array and call it for each image, instead of writing the same code multiple times. 您应该编写一种将位图加载到double[900]
数组中并为每个图像调用它的方法,而不是多次编写相同的代码。 This helps to reduce problems like this, where different results are given by two pieces of code that should return the same result. 这有助于减少类似的问题,在该问题中,应返回相同结果的两段代码给出了不同的结果。
I tried your code. 我尝试了您的代码。 It helped me too and thanks a lot for that. 它也对我有帮助,对此非常感谢。 I could get your code working by doing some changes to getting bit array from the image. 通过对从图像中获取位数组进行一些更改,可以使您的代码正常工作。 Here's the method I used. 这是我使用的方法。
`
private double[] GetImageData(Bitmap bmp)
{
double[] imageData = null;
//Make the image grayscale
Grayscale filter = new Grayscale(0.2125, 0.7154, 0.0721);
bmp = filter.Apply(bmp);
//Binarize the image
AForge.Imaging.Filters.Threshold thFilter = new AForge.Imaging.Filters.Threshold(128);
thFilter.ApplyInPlace(bmp);
int height = bmp.Height;
int width = bmp.Width;
imageData = new double[height * width];
int imagePointer = 0;
System.Diagnostics.Debug.WriteLine("Height : " + height);
System.Diagnostics.Debug.WriteLine("Width : " + width);
for (int i = 0; i < height; i++)
{
for (int j = 0; j < width; j++)
{
System.Diagnostics.Debug.Write(string.Format("({0} , {1}) Color : {2}\n", i, j, bmp.GetPixel(i, j)));
//Identify the black points of the image
if (bmp.GetPixel(i, j) == Color.FromArgb(255, 0, 0, 0))
{
imageData[imagePointer] = 1;
}
else
{
imageData[imagePointer] = 0;
}
imagePointer++;
}
System.Diagnostics.Debug.WriteLine("");
}
System.Diagnostics.Debug.WriteLine("Bits : " + imagePointer );
return imageData;
}`
Hope this will help. 希望这会有所帮助。 Thanks. 谢谢。
try this 尝试这个
double error = learning.RunEpoch(input, new double[9][] { new double[3] **{ 1, -1, -1 }**,new double[3] **{ -1, 1, -1 }**,new double[3] **{ -1, -1, 1 }**,//A
new double[3] **{ 1, -1, -1 }**,new double[3] **{ -1, 1, -1 }**,new double[3] **{ -1, -1, 1 }**,//B
new double[3] **{ 1, -1, -1 }**,new double[3] **{ -1, 1, -1 }**,new double[3] **{ -1, -1, 1 }** }//C
);
or this way 或者这样
double[][] output = new double[patterns][];
for (int j = 0; j < patterns; j++)
{
output[j] = new double[patterns];
for (int i = 0; i < patterns; i++)
{
if (i != j)
{
output[j][i] = -1;
}
else
{
output[j][i] = 1;
}
}
}
double error = learning.RunEpoch(input,output)
double[] netout = neuralNet.Compute(pattern);
int maxIndex = 0;
double max = netout[0];
for (int i = 1; i < netout.Length; i++)
{
if (netout[i] > max)
{
max = netout[i];
maxIndex = i;
}
}
if maxIndex=0 answer is A 如果maxIndex = 0答案是A
if maxIndex=1 answer is B 如果maxIndex = 1答案是B
if maxIndex=2 answer is C 如果maxIndex = 2答案是C
also I think that you must create matrix from images and use it as pattern, for example 20/20 or 15/15 or small, your 30/30 is big. 我还认为您必须根据图像创建矩阵并将其用作图案,例如20/20或15/15或更小,则30/30很大。
I use different way for get Image Scheme. 我使用不同的方式来获取图像方案。 I Divide image 20/20 and If one of pixels in rectangle is black (or another colour you want) save 1 in matrix, else 0. 我将图像除以20/20,如果矩形中的一个像素为黑色(或您想要的另一种颜色),则在矩阵中保存1,否则为0。
I make replace all pixels an after this I have only two colours, White and Black, I can manipulate with contour. 我将所有像素全部替换掉,之后只有白色和黑色两种颜色,可以按轮廓操作。
private void Cmd_ReplaceColors(ref WriteableBitmap Par_WriteableBitmap,int Par_Limit=180)
{
for (int y = 0; y < Par_WriteableBitmap.PixelHeight; y++)
{
for (int x = 0; x < Par_WriteableBitmap.PixelWidth; x++)
{
Color color = Par_WriteableBitmap.GetPixel(x, y);
if (color == Colors.White)
{
}
else
{
if (color.R < Par_Limit)
{
Par_WriteableBitmap.SetPixel(x, y, Colors.Black);
}
else
{
Par_WriteableBitmap.SetPixel(x, y, Colors.White);
}
}
}
}
Par_WriteableBitmap.Invalidate();
}
1000 iterations in my opinion is small, better 100 000 :) 我认为1000次迭代很小,最好100 000次:)
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.