自定义过滤器组未生成预期输出

Question

我已经实现了 4.1 (Pre-processing).

部分

The preprocessing step aims to enhance image features along a set of chosen directions. First, image is grey-scaled and filtered with a sharpening filter (we subtract from the image its local-mean filtered version), thus eliminating the DC component.

We selected 12 not overlapping filters, to analyze 12 different directions, rotated with respect to 15° each other.

GitHub Repositiry is here.

由于文章中给出的公式不正确，我尝试了两组不同的公式。

第一组公式，

第二组公式，

预期的输出应该是，

他们都没有给出正确的结果。

有人可以建议我修改吗？

GitHub Repository is here.

最相关的源代码部分在这里：

    public List<Bitmap> Apply(Bitmap bitmap)
    {
        Kernels = new List<KassWitkinKernel>();

        double degrees = FilterAngle;

        KassWitkinKernel kernel;
        for (int i = 0; i < NoOfFilters; i++)
        {
            kernel = new KassWitkinKernel();
            kernel.Width = KernelDimension;
            kernel.Height = KernelDimension;
            kernel.CenterX = (kernel.Width) / 2;
            kernel.CenterY = (kernel.Height) / 2;
            kernel.Du = 2;
            kernel.Dv = 2;
            kernel.ThetaInRadian = Tools.DegreeToRadian(degrees);
            kernel.Compute();

            //SleuthEye
            kernel.Pad(kernel.Width, kernel.Height, WidthWithPadding, HeightWithPadding);

            Kernels.Add(kernel);

            degrees += degrees;
        }

        List<Bitmap> list = new List<Bitmap>();

        Bitmap image = (Bitmap)bitmap.Clone();

        //PictureBoxForm f = new PictureBoxForm(image);
        //f.ShowDialog();

        Complex[,] cImagePadded = ImageDataConverter.ToComplex(image);

        Complex[,] fftImage = FourierTransform.ForwardFFT(cImagePadded);

        foreach (KassWitkinKernel k in Kernels)
        {
            Complex[,] cKernelPadded = k.ToComplexPadded();
            Complex[,] convolved = Convolution.ConvolveInFrequencyDomain(fftImage, cKernelPadded);

            Bitmap temp = ImageDataConverter.ToBitmap(convolved);



            list.Add(temp);
        }

        return list;
    }

Answer 1

也许首先应该提到的是应该以 FilterAngle（在您的情况下为 15 度）增量中增加的角度生成过滤器。这可以通过如下修改 KassWitkinFilterBank.Apply 来完成（参见 this commit）：

public List<Bitmap> Apply(Bitmap bitmap)
{
    // ...

    // The generated template filter from the equations gives a line at 45 degrees. 
    // To get the filter to highlight lines starting with an angle of 90 degrees
    // we should start with an additional 45 degrees offset.
    double degrees = 45;

    KassWitkinKernel kernel;
    for (int i = 0; i < NoOfFilters; i++)
    {
        // ... setup filter (unchanged)

        // Now increment the angle by FilterAngle
        // (not "+= degrees" which doubles the value at each step)
        degrees += FilterAngle;
    }

这应该会给您以下结果：

这不完全是论文的结果，图像之间的差异仍然非常细微，但您应该能够注意到第 8 个数字中的划痕线最强烈（正如自划痕角度约为 100-105 度）。

为了改善结果，我们应该按照论文中描述的相同方式为过滤器提供预处理图像：

First, image is grey-scaled and filtered with a sharpening filter (we subtract from the image its local-mean filtered version), thus eliminating the DC component

当你这样做时，你会得到一个值矩阵，其中一些是负数。因此，这个中间处理结果不适合存储为 Bitmap。作为图像处理的一般规则，您应该根据需要将所有中间结果保留在 double 或 Complex 中，并且只将最终结果转换回 Bitmap 以进行可视化。

整合您的更改以添加来自 your GitHub repository while keeping intermediate results as doubles can be achieve by changing the input bitmap and temporary image variables to use double[,] datatype instead of Bitmap in the KassWitkinFilterBank.Apply method (see this commit 的图像锐化):

public List<Bitmap> Apply(double[,] bitmap)
{
    // [...]

    double[,] image = (double[,])bitmap.Clone();

    // [...]
}

应该会得到以下结果：

或者为了更好地突出区别，左边是图1（0度），右边是图8（105度）：

自定义过滤器组未生成预期输出

Custom filter bank is not generating the expected output

c#-4.0

image-processing

convolution

bitmapdata

imagefilter