Table of Contents

Geometric Transformations

This example introduces OpenCV operations that change an image's size, orientation, or coordinate system. OpenCvSharp follows the native OpenCV model: a source Mat, a destination Mat, a transformation definition, and an interpolation or border policy.

Resize an image

Specify the final dimensions when the output size is known:

using var source = Cv2.ImRead("input.jpg", ImreadModes.Color);
if (source.Empty())
{
    throw new IOException("Could not read input.jpg.");
}

using var resized = new Mat();
Cv2.Resize(
    source,
    resized,
    new Size(640, 480),
    interpolation: InterpolationFlags.Area);

InterpolationFlags.Area is normally a good choice when reducing an image. InterpolationFlags.Linear is a common default for enlargement, while InterpolationFlags.Nearest preserves exact class values in label images and masks.

To preserve the aspect ratio, calculate one dimension:

const int targetWidth = 800;
double scale = (double)targetWidth / source.Width;
int targetHeight = (int)Math.Round(source.Height * scale);

using var resized = new Mat();
Cv2.Resize(
    source,
    resized,
    new Size(targetWidth, targetHeight),
    interpolation: InterpolationFlags.Area);

Crop with a region of interest

A region of interest is a lightweight Mat header that views the source image data:

var roi = new Rect(x: 100, y: 80, width: 320, height: 240);
using Mat view = source.SubMat(roi);

The rectangle must remain inside the source image. The view and source share pixel storage, so modifications through view affect source, and source must remain alive while the view is used.

Clone the view when the cropped image must own independent storage or outlive the source:

using Mat cropped = view.Clone();

Rotate by a right angle

Use Cv2.Rotate for exact 90-degree operations:

using var rotated = new Mat();
Cv2.Rotate(source, rotated, RotateFlags.Rotate90Clockwise);

This is simpler than constructing an affine matrix and does not interpolate pixels.

Rotate by an arbitrary angle

Create a 2-by-3 affine transformation matrix and pass it to Cv2.WarpAffine:

var center = new Point2f(source.Width / 2f, source.Height / 2f);
using Mat rotation = Cv2.GetRotationMatrix2D(center, angle: 15, scale: 1);
using var rotated = new Mat();

Cv2.WarpAffine(
    source,
    rotated,
    rotation,
    source.Size(),
    InterpolationFlags.Linear,
    BorderTypes.Constant,
    new Scalar(0, 0, 0));

The destination size is independent of the transformation. Keeping source.Size() can clip corners after rotation. Calculate a larger destination and adjust the translation terms in the matrix when the entire rotated image must remain visible.

Rectify a quadrilateral

A perspective transformation maps four source points to four destination points. The point order must correspond between the two arrays:

Point2f[] sourceCorners =
[
    new(120, 80),
    new(920, 110),
    new(880, 680),
    new(150, 650),
];

Point2f[] destinationCorners =
[
    new(0, 0),
    new(799, 0),
    new(799, 599),
    new(0, 599),
];

using Mat transform = Cv2.GetPerspectiveTransform(
    sourceCorners,
    destinationCorners);
using var rectified = new Mat();

Cv2.WarpPerspective(
    source,
    rectified,
    transform,
    new Size(800, 600));

Use a consistent order such as top-left, top-right, bottom-right, bottom-left. A crossed or inconsistent order produces a folded or mirrored result.

Interpolation and border choices

The official OpenCV reference defines the exact interpolation formulas and supported flags. In typical OpenCvSharp applications:

  • Use Area when shrinking photographs.
  • Use Linear for general-purpose resizing and warping.
  • Use Cubic or Lanczos4 when enlargement quality matters more than speed.
  • Use Nearest for masks, labels, and other discrete values.
  • Use BorderTypes.Constant when pixels outside the image should have a known value.
  • Use BorderTypes.Replicate when a warp should extend the nearest edge pixels.
  • Use reflection modes such as BorderTypes.Reflect101 only with APIs whose OpenCV reference explicitly supports them, commonly filter operations.

Always pass destination dimensions as (width, height) through Size. Image indexing and Mat construction otherwise commonly use (rows, columns), which correspond to (height, width).

From C++ or Python

The native functions cv::resize, cv::warpAffine, and cv::warpPerspective map to Cv2.Resize, Cv2.WarpAffine, and Cv2.WarpPerspective. Python commonly returns a new NumPy array, while OpenCvSharp normally receives a destination Mat that the caller owns and disposes.

Official OpenCV references

Edit this page