Intro to computer vision in .net

@slorello
Intro to Computer Vision in .NET
Steve Lorello
.NET Developer advocate @Vonage
Twitter: @slorello

@slorello
What is Computer Vision?

@slorello
“ “The Goal of computer vision
is to write computer
programs that can interpret
images
Steve Seitz

@slorello
1. What is a Digital Image?
2. Hello OpenCV in .NET
3. Convolution and Edge Detection
4. Facial Detection
5. Facial Detection with Vonage Video API
6. Feature Tracking and Image Projection
Agenda

@slorello
What is a
Digital
Image?

@slorello
● An Image is a Function
● A function of Intensity Values
at Given Positions
● Those Intensity Values Fall
Along an Arbitrary Range

@slorello Source: Aaron Bobick’s Intro to Computer Vision Udacity

@slorello
Using Computer Vision in .NET

@slorello
● OpenCV (Open Source Computer Vision
Library): https://opencv.org/
● Emgu CV: http://www.emgu.com/

@slorello
● Create a Project in Visual Studio
● Install EmguCv with package manager:
Emgu.CV.runtime.<platform>

@slorello https://github.com/slorello89/ShowImage
var zero = CvInvoke.Imread(Path.Join("resources","zero.jpg"));
CvInvoke.Imshow("zero", zero);
CvInvoke.WaitKey(0);

@slorello
Convolution and Edge Detection

@slorello https://carbon.now.sh/

@slorello http://homepages.inf.ed.ac.uk/rbf/HIPR2/sobel.htm
Sobel Operator

@slorello https://github.com/slorello89/BasicSobel
CvInvoke.CvtColor(img, gray, Emgu.CV.CvEnum.ColorConversion.Bgr2Gray);
CvInvoke.GaussianBlur(gray, gray, new System.Drawing.Size(3, 3), 0);
CvInvoke.Sobel(gray, gradX, Emgu.CV.CvEnum.DepthType.Cv16S, 1, 0, 3);
CvInvoke.Sobel(gray, gradY, Emgu.CV.CvEnum.DepthType.Cv16S, 0, 1, 3);
CvInvoke.ConvertScaleAbs(gradX, absGradX, 1, 0);
CvInvoke.ConvertScaleAbs(gradY, absGradY, 1, 0);
CvInvoke.AddWeighted(absGradX, .5, absGradY, .5, 0, sobelGrad);

@slorello
Gradient in X Gradient in Y

@slorello Source: https://dsp.stackexchange.com/
Gaussian Kernel

@slorello Source: https://www.globalsino.com/EM/page1371.html
Sharpening Filter

@slorello https://github.com/slorello89/Convolution
//blur
CvInvoke.GaussianBlur(zero, blurred, new System.Drawing.Size(9, 9), 9);
var blurredImage = blurred.ToImage<Bgr, byte>();
//sharpen
var detail = (zero - blurredImage) * 2;
var sharpened = zero + detail;

@slorello
Here it is at 10X detail

@slorello
1. Use Haar-Like features as masks
2. Use integral images to calculate relative
shading per these masks
3. Use a Cascading Classiﬁer to detect faces
Viola-Jones Technique

@slorello https://www.quora.com/How-can-I-understand-Haar-like-feature-for-face-detection
Haar-like features

@slorello Source https://www.mathworks.com/help/images/integral-image.html
Integral Images or Summed Area table

@slorello
● Construct Cascading Classifier
● Run Classification
● Use Rectangles from classification to draw
boxes around faces

@slorello https://github.com/slorello89/FacialDetection
var faceClassifier = new CascadeClassifier(Path.Join("resources",
"haarcascade_frontalface_default.xml"));
var img = CvInvoke.Imread(Path.Join("resources", "imageWithFace.jpg"));
var faces = faceClassifier.DetectMultiScale(img,
minSize: new System.Drawing.Size(300,300));
foreach(var face in faces)
{
CvInvoke.Rectangle(img, face,
new Emgu.CV.Structure.MCvScalar(255, 0, 0), 10);
}

@slorello
Face Detection With the Vonage Video API
https://www.vonage.com/communications-apis/video/

@slorello
● Create a WPF app
● Add the OpenTok.Client SDK to it
● Add a new class implementing IVideoRender
called and extending Control
FaceDetectionVideoRenderer
● Add a Control to the Main Xaml ﬁle where we’ll
put publisher video - call it “PublisherVideo”
● Add a Detect Faces and Connect button

@slorello https://github.com/opentok-community/wpf-facial-detection
Publisher = new Publisher(Context.Instance,
renderer: PublisherVideo);
Session = new Session(Context.Instance, API_KEY, SESSION_ID);

private void Connect_Click(object sender, RoutedEventArgs e)
{
if (Disconnect)
{
Session.Unpublish(Publisher);
Session.Disconnect();
}
else
{
Session.Connect(TOKEN);
}
Disconnect = !Disconnect;
ConnectDisconnectButton.Content = Disconnect ? "Disconnect" : "Connect";
}

private void DetectFacesButton_Click(object sender, RoutedEventArgs e)
{
PublisherVideo.ToggleFaceDetection(!PublisherVideo.DetectingFaces);
foreach (var subscriber in SubscriberByStream.Values)
{
((FaceDetectionVideoRenderer)subscriber.VideoRenderer)
.ToggleFaceDetection(PublisherVideo.DetectingFaces);
}
}

private void Session_StreamReceived(object sender, Session.StreamEventArgs e)
{
FaceDetectionVideoRenderer renderer = new FaceDetectionVideoRenderer();
renderer.ToggleFaceDetection(PublisherVideo.DetectingFaces);
SubscriberGrid.Children.Add(renderer);
UpdateGridSize(SubscriberGrid.Children.Count);
Subscriber subscriber = new Subscriber(Context.Instance, e.Stream, renderer);
SubscriberByStream.Add(e.Stream, subscriber);
Session.Subscribe(subscriber);
}

@slorello
● Intercept each frame before it’s rendered.
● Run face detection on each frame
● Draw a rectangle on each frame to show
where the face is
● Render the Frame

VideoBitmap = new WriteableBitmap(frame.Width,
frame.Height, 96, 96, PixelFormats.Bgr32, null);
if (Background is ImageBrush)
{
ImageBrush b = (ImageBrush)Background;
b.ImageSource = VideoBitmap;
}

@slorello https://romannurik.github.io/SlidesCodeHighlighter/
if (VideoBitmap != null)
{
VideoBitmap.Lock();
IntPtr[] buffer = { VideoBitmap.BackBuffer };
int[] stride = { VideoBitmap.BackBufferStride };
frame.ConvertInPlace(OpenTok.PixelFormat.FormatArgb32, buffer, stride);
if (DetectingFaces)
{
using (var image = new Image<Bgr, byte>(frame.Width, frame.Height, stride[0], buffer[0]))
{
if (_watch.ElapsedMilliseconds > INTERVAL)
{
var reduced = image.Resize(1.0 / SCALE_FACTOR, Emgu.CV.CvEnum.Inter.Linear);
_watch.Restart();
_images.Add(reduced);
}
}
DrawRectanglesOnBitmap(VideoBitmap, _faces);
}
VideoBitmap.AddDirtyRect(new Int32Rect(0, 0, FrameWidth, FrameHeight));
VideoBitmap.Unlock();
}

System.Threading.ThreadPool.QueueUserWorkItem(delegate
{
try
{
while (true)
{
using (var image = _images.Take(token))
{
_faces = _profileClassifier.DetectMultiScale(image);
}
}
}
catch (OperationCanceledException)
{
//exit gracefully
}
}, null);

public static void DrawRectanglesOnBitmap(WriteableBitmap bitmap, Rectangle[] rectangles)
{
foreach (var rect in rectangles)
{
var x1 = (int)((rect.X * (int)SCALE_FACTOR) * PIXEL_POINT_CONVERSION);
var x2 = (int)(x1 + (((int)SCALE_FACTOR * rect.Width) * PIXEL_POINT_CONVERSION));
var y1 = rect.Y * (int)SCALE_FACTOR;
var y2 = y1 + ((int)SCALE_FACTOR * rect.Height);
bitmap.DrawLineAa(x1, y1, x2, y1, strokeThickness: 5, color: Colors.Blue);
}
}

@slorello
Feature Detection, Tracking, Image
Projection
https://www.vonage.com/communications-apis/video/

@slorello
● What’s a good Feature?
● Detect Features with Orb
● Feature Tracking with a BF
tracker
● Project an image.

@slorello
● A good feature is a part
of the image, where
there are multiple edges
● Thus we often think of
them as Corners
● We can use the ORB
method (Oriented FAST
and rotated BRIEF)
https://www.slideshare.net/slksaad/multiimage-matching-using-multiscale
-oriented-patches

@slorello https://github.com/slorello89/FeatureDetection
var orbDetector = new ORBDetector(10000);
var features1 = new VectorOfKeyPoint();
var descriptors1 = new Mat();
orbDetector.DetectAndCompute(img, null, features1, descriptors1, false);
Features2DToolbox.DrawKeypoints(img, features1, img, new Bgr(255, 0, 0));

@slorello
● Now that we have some features we can
match them to features in other images!
● We’ll use K-nearest-neighbors matching
on the Brute-force matcher

var bfMatcher = new BFMatcher(DistanceType.L1);
bfMatcher.Add(descriptors1);
bfMatcher.KnnMatch(descriptors2, knnMatches, k:1,mask:null,compactResult:true);
foreach(var matchSet in knnMatches.ToArrayOfArray())
{
if(matchSet.Length>0 && matchSet[0].Distance < 400)
{
matchList.Add(matchSet[0]);
var featureModel = features1[matchSet[0].TrainIdx];
var featureTrain = features2[matchSet[0].QueryIdx];
srcPts.Add(featureModel.Point);
dstPts.Add(featureTrain.Point);
}
}
var matches = new VectorOfDMatch(matchList.ToArray());
var imgOut = new Mat();
Features2DToolbox.DrawMatches(img, features1, img2, features2, matches,
imgOut, new MCvScalar(255, 0, 0), new MCvScalar(0, 0, 255));

@slorello
● Image transformations
● 8 degrees of freedom
● Need at least 4 matches
● Homographies
Image Projection

@slorello https://inst.eecs.berkeley.edu/~cs194-26/fa17/upload/ﬁles/proj6B/cs194-26-aap/h2.png

var srcPoints = InputImageToPointCorners(cat);
var dstPoints = FaceToCorners(face);
var homography = CvInvoke.FindHomography(srcPoints, dstPoints,
Emgu.CV.CvEnum.RobustEstimationAlgorithm.Ransac, 5.0);
CvInvoke.WarpPerspective(cat, projected, homography, img.Size);
img.Mat.CopyTo(projected, 1 - projected);

@slorello
A Little More About Me
● .NET Developer & Software Engineer
● .NET Developer Advocate @Vonage
● Computer Science Graduate Student
@GeorgiaTech - specializing in Computer
Perception
● Blog posts: https://dev.to/slorello or
https://www.nexmo.com/blog/author/stevelorello
● Twitter: @slorello

@slorello
https://github.com/slorello89/ShowImage
https://github.com/opentok-community/wpf-facial-detection
https://github.com/slorello89/BasicSobel
https://github.com/slorello89/FacialDetection
http://www.emgu.com/
https://opencv.org/
https://tokbox.com/developer/tutorials/
https://developer.nexmo.com/
https://www.nexmo.com/blog/2020/03/18/real-time-face-detec
tion-in-net-with-opentok-and-opencv-dr
Resources
LinkedIn: https://www.linkedin.com/in/stephen-lorello-143086a9/
Twitter: @slorello

@slorello Attribution if needed

@slorello
An image
with some
text on the
side.
URL ATTRIBUTION GOES HERE

@slorello
An image with some text over it
Attribution if needed

@slorello
“ “A really large quote would
go here so everyone can
read it.
Some Persons Name
https://website.com

@slorello
Code Snippet Examples

@slorello https://romannurik.github.io/SlidesCodeHighlighter/
var faceClassifier = new CascadeClassifier(Path.Join("resources",
"haarcascade_frontalface_default.xml"));
var img = CvInvoke.Imread(Path.Join("resources", "imageWithFace.jpg"));
var faces = faceClassifier.DetectMultiScale(img,
minSize: new System.Drawing.Size(300,300));
foreach(var face in faces)
{
CvInvoke.Rectangle(img, face,
new Emgu.CV.Structure.MCvScalar(255, 0, 0), 10);
}

@slorello
Example Web Page Slides

@slorello https://developer.nexmo.com

@slorello https://developer.nexmo.com
Website in a
mobile phone.

Intro to computer vision in .net

Recommended

More Related Content

What's hot (18)

Similar to Intro to computer vision in .net (20)

Recently uploaded (20)

Intro to computer vision in .net