检测图像中的正方形

Detecting square in image

在一所学校,我们正在准备扫描过的图稿并希望自动裁剪为正确的尺寸。孩子们(尝试)在一个矩形内画画:

我想检测内部矩形边框,所以我应用了一些过滤器 accord.net:

var newImage = new Bitmap(@"C:\Temp\temp.jpg");
var g = Graphics.FromImage(newImage);
var pen = new Pen(Color.Purple, 10);

var grayScaleFilter = new Grayscale(1, 0, 0);
var image = grayScaleFilter.Apply(newImage);
image.Save(@"C:\temp\grey.jpg");

var skewChecker = new DocumentSkewChecker();
var angle = skewChecker.GetSkewAngle(image);
var rotationFilter = new RotateBilinear(-angle);
rotationFilter.FillColor = Color.White;
var rotatedImage = rotationFilter.Apply(image);
rotatedImage.Save(@"C:\Temp\rotated.jpg");

var thresholdFilter = new IterativeThreshold(10, 128);
thresholdFilter.ApplyInPlace(rotatedImage);
rotatedImage.Save(@"C:\temp\threshold.jpg");

var invertFilter = new Invert();
invertFilter.ApplyInPlace(rotatedImage);
rotatedImage.Save(@"C:\temp\inverted.jpg");

var bc = new BlobCounter
{
    BackgroundThreshold = Color.Black,
    FilterBlobs = true,
    MinWidth = 1000,
    MinHeight = 1000
};

bc.ProcessImage(rotatedImage);
foreach (var rect in bc.GetObjectsRectangles())
{
    g.DrawRectangle(pen, rect);
}

newImage.Save(@"C:\Temp\test.jpg");

这会生成以下倒置图像,BlobCounter 将其用作输入:

但是 blobcounter 的结果不是非常准确,紫色线表示 BC 检测到的内容。

accord.net 中的 BlobCounter 是否有更好的替代方案,或者是否有其他 C# 库更适合这种计算机视觉?

这是我午休无聊时的一个简单解决方案。

基本上它只是针对给定的颜色阈值(黑色)从外到内扫描所有维度,然后取最突出的结果。

给定

[MethodImpl(MethodImplOptions.AggressiveInlining)]
public static unsafe bool IsValid(int* scan0Ptr, int x, int y,int stride, double thresh)
{
   var c = *(scan0Ptr + x + y * stride);
   var r = ((c >> 16) & 255);
   var g = ((c >> 8) & 255);
   var b = ((c >> 0) & 255);

   // compare it against the threshold
   return r * r + g * g + b * b < thresh;
}

private static int GetBest(IEnumerable<int> array)
   => array.Where(x => x != 0)
      .GroupBy(i => i)
      .OrderByDescending(grp => grp.Count())
      .Select(grp => grp.Key)
      .First();

例子

private static unsafe Rectangle ConvertImage(string path, Color source,  double threshold)
{

   var thresh = threshold * threshold;

   using var bmp = new Bitmap(path);

   // lock the array for direct access
   var bitmapData = bmp.LockBits(new Rectangle(0, 0, bmp.Width, bmp.Height), ImageLockMode.ReadWrite, PixelFormat.Format32bppPArgb);
   int left, top, bottom, right;

   try
   {
      // get the pointer
      var scan0Ptr = (int*)bitmapData.Scan0;    
      // get the stride
      var stride = bitmapData.Stride / 4;

      var array = new int[bmp.Height];

      for (var y = 0; y < bmp.Height; y++)
      for (var x = 0; x < bmp.Width; x++)
         if (IsValid(scan0Ptr, x, y, stride, thresh))
         {
            array[y] = x;
            break;
         }

      left = GetBest(array);

      array = new int[bmp.Height];

      for (var y = 0; y < bmp.Height; y++)
      for (var x = bmp.Width-1; x > 0; x--)
         if (IsValid(scan0Ptr, x, y, stride, thresh))
         {
            array[y] = x;
            break;
         }

      right = GetBest(array);

      array = new int[bmp.Width];

      for (var x = 0; x < bmp.Width; x++)
      for (var y = 0; y < bmp.Height; y++)
         if (IsValid(scan0Ptr, x, y, stride, thresh))
         {
            array[x] = y;
            break;
         }

      top = GetBest(array);

      array = new int[bmp.Width];

      for (var x = 0; x < bmp.Width; x++)
      for (var y = bmp.Height-1; y > 0; y--)
         if (IsValid(scan0Ptr, x, y, stride, thresh))
         {
            array[x] = y;
            break;
         }

      bottom = GetBest(array);


   }
   finally
   {
      // unlock the bitmap
      bmp.UnlockBits(bitmapData);
   }

   return new Rectangle(left,top,right-left,bottom-top);

}

用法

var fileName = @"D:48p.jpg";

var rect = ConvertImage(fileName, Color.Black, 50);

using var src = new Bitmap(fileName);
using var target = new Bitmap(rect.Width, rect.Height);
using var g = Graphics.FromImage(target);

g.DrawImage(src, new Rectangle(0, 0, target.Width, target.Height), rect, GraphicsUnit.Pixel); 

target.Save(@"D:\Test.Bmp");

输出

备注:

  • 这并不意味着万无一失或最佳解决方案。只是一个快速简单的。
  • 有很多方法可以解决这个问题,即使是机器学习方法也可能更好、更稳健。
  • 这里有很多代码重复,基本上我只是对每一面进行了复制、粘贴和调整
  • 我刚刚选择了一个似乎有效的任意阈值。玩一玩
  • 获得边最常见的情况可能不是最好的方法,也许你会想要对结果进行分类。
  • 你或许可以理智地限制一侧需要扫描的数量。