Table of Contents

Build an Image Processing Pipeline

Most classical computer vision programs follow a pipeline: load an image, normalize it into the representation an algorithm expects, isolate interesting pixels, extract objects, and produce a result. This example detects sufficiently large bright regions and draws their contours.

Complete example

using OpenCvSharp;

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

using var grayscale = new Mat();
Cv2.CvtColor(source, grayscale, ColorConversionCodes.BGR2GRAY);

using var blurred = new Mat();
Cv2.GaussianBlur(grayscale, blurred, new Size(5, 5), sigmaX: 0);

using var binary = new Mat();
Cv2.Threshold(
    blurred,
    binary,
    thresh: 0,
    maxval: 255,
    type: ThresholdTypes.Binary | ThresholdTypes.Otsu);

using var kernel = Cv2.GetStructuringElement(MorphShapes.Rect, new Size(3, 3));
using var cleaned = new Mat();
Cv2.MorphologyEx(binary, cleaned, MorphTypes.Open, kernel);

Cv2.FindContours(
    cleaned,
    out Point[][] contours,
    out _,
    RetrievalModes.External,
    ContourApproximationModes.ApproxSimple);

Point[][] significantContours = contours
    .Where(contour => Cv2.ContourArea(contour) >= 100)
    .ToArray();

using var annotated = source.Clone();
Cv2.DrawContours(
    annotated,
    significantContours,
    contourIdx: -1,
    color: Scalar.Red,
    thickness: 2);

Cv2.ImWrite("result.png", annotated);

What each stage does

Normalize the input

CvtColor produces a single-channel image because thresholding intensity is easier to reason about than thresholding three BGR channels independently. Choose the color conversion that matches the meaning of the input; BGR2GRAY is only one possibility.

Reduce noise

GaussianBlur suppresses small intensity variations before thresholding. Kernel size and sigma affect how much detail is removed. Avoid copying a kernel size from an example without checking it against the size of objects in the real images.

Create a binary mask

The combination of Binary and Otsu asks OpenCV to estimate one global threshold from the image histogram. It is a good baseline for images with two well-separated intensity groups. Uneven illumination may require AdaptiveThreshold, illumination correction, or a different segmentation method.

Clean the mask

Morphological opening removes small foreground spots. Closing instead fills small holes and gaps. The structuring-element shape and size encode what counts as "small," so they should be treated as model parameters rather than universal constants.

Extract and filter objects

FindContours returns arrays of points around connected shapes. RetrievalModes.External ignores nested contours, which is appropriate when only outer boundaries matter. Filter by area, geometry, or domain-specific measurements before drawing or analyzing the result.

Debug a pipeline stage by stage

Save intermediate images while tuning a pipeline:

Cv2.ImWrite("debug-01-grayscale.png", grayscale);
Cv2.ImWrite("debug-02-binary.png", binary);
Cv2.ImWrite("debug-03-cleaned.png", cleaned);

When a final result is wrong, intermediate images show which stage first lost the desired information. This is usually more useful than tuning several parameters at once.

Official OpenCV references

Edit this page