OCR与Aforge.net的感知神经网络回答错误

Question

我试图用C#中的Aforge.Net通过感知器进行OCR.我用九个30*30二进制图片学习了我的网络.但在结果中,它将所有内容都识别为"C".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}
    }

我真的很想知道它为什么回答错误.

Answer 1

将初始 30x30 图像加载到结构中的 double[900] 数组中时，input您使用以下计算：

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;
    }

您的偏移量计算在这里是错误的。您需要更改j * 10 + k为j * 30 + k，否则您将得到无效结果。稍后，您在加载测试图像时使用正确的偏移计算，这就是它无法与损坏的样本正确匹配的原因。

您应该编写一个方法将位图加载到double[900]数组中并为每个图像调用它，而不是多次编写相同的代码。这有助于减少此类问题，即两段应该返回相同结果的代码给出不同的结果。