IList <PreProcessedImage> ITrainModel.startTrainingModel()
{
IList <PreProcessedImage> result = new List <PreProcessedImage>(this.trainingModel.Sum(item => item.listOfImages.Count));
using (PreProcess preProcesAlgorithm = new PreProcess(null))
{
int count = 0;
foreach (ImageCategory imageCategory in this.trainingModel)
{
Console.WriteLine(count++ + " Training category: " + imageCategory.leafCategory + "total files: " + imageCategory.listOfImages.Count);
foreach (string imagePath in imageCategory.listOfImages)
{
PreProcessedImage preProcessedImage = preProcesAlgorithm.Execute(imagePath, imageCategory.leafCategory);
result.Add(preProcessedImage);
}
}
}
return(result);
}
public PreProcessedImage Execute(string imagePath, string category)
{
PreProcessedImage result = new PreProcessedImage();
result.FilePath = imagePath;
result.Category = category;
using (PreProcess preProcessAlgorithm = new PreProcess(new Image <Bgr, Byte>(imagePath)))
using (Image <Gray, Byte> grayScaleImage = preProcessAlgorithm._ImageToGrayScaleUsingConvert())
{
using (FeatureExtractAlgorithm featureSet = new FeatureExtractAlgorithm(grayScaleImage))
using (Mat canny = new Mat())
using (VectorOfVectorOfPoint contours = new VectorOfVectorOfPoint())
{
KeyPoints descriptor = featureSet.SIFTDescriptor();
result.KeyPoints = descriptor;
result.ContourArea = 0;
CvInvoke.Canny(grayScaleImage, canny, 100, 50);
int[,] hierarchy = CvInvoke.FindContourTree(canny, contours, ChainApproxMethod.ChainApproxSimple);
result.NumberOfContours = contours.Size;
double maxArea = double.MinValue;
int maxIndex = -1;
for (int index = 0; index < contours.Size; index++)
{
double area = CvInvoke.ContourArea(contours[index]);
result.ContourArea += area;
if (area > maxArea)
{
maxArea = area;
maxIndex = index;
}
}
result.Contour = new VectorOfPoint();
CvInvoke.ApproxPolyDP(contours[maxIndex], result.Contour, CvInvoke.ArcLength(contours[maxIndex], true) * 0.02, true);
return(result);
}
}
}
private List <MatcherResult> QueryImage(string filePath, string expectedCategory)
{
using (PreProcess preProcessAlgorithm = new PreProcess(null))
{
PreProcessedImage preProcessedQueryImage = preProcessAlgorithm.Execute(filePath, expectedCategory);
List <MatcherResult> finalResultSet = new List <MatcherResult>(this.trainingDataset.Count);
DescriptorMatcher learningAlgo = new DescriptorMatcher();
foreach (PreProcessedImage trainingData in this.trainingDataset)
{
try
{
using (BFMatcher trainingMatcher = new BFMatcher(DistanceType.L1))
{
double areaRatio = trainingData.ContourArea / preProcessedQueryImage.ContourArea;
int contourDelta = trainingData.NumberOfContours - preProcessedQueryImage.NumberOfContours;
//if (areaRatio < 0.1 || areaRatio > 5)
//{
// continue;
//}
//if(contourDelta > 100 || contourDelta < -100)
//{
// continue;
//}
trainingMatcher.Add(trainingData.KeyPoints.Descriptor);
MatcherResult result = learningAlgo.knnMatch(preProcessedQueryImage.KeyPoints, trainingMatcher, trainingData.Category, 300, trainingData);
result.ContourRatio = CvInvoke.MatchShapes(trainingData.Contour, preProcessedQueryImage.Contour, Emgu.CV.CvEnum.ContoursMatchType.I3);
result.AreaRatio = areaRatio;
result.ContoursDelta = contourDelta;
//Console.WriteLine("QueryCategory: {0}, {1}", preProcessedQueryImage.Category, result.ToString());
finalResultSet.Add(result);
}
}
catch (Exception exception)
{
//Console.WriteLine(exception);
}
}
//IEnumerable<MatcherResult> highConfidenceResults = finalResultSet.Where(item => item.MatchingPoints >= 3);
//finalResultSet = highConfidenceResults.ToList();
//// If there are clusters with size great than or equal to 3 , then they have a very high probability of being correct.
//if (highConfidenceResults.Any())
//{
// //Console.WriteLine("High confidence:" + highConfidenceResults.Count() + " : " + highConfidenceResults.Max(item => item.MatchingPoints));
// finalResultSet = highConfidenceResults.ToList();
//}
//else
//{
// finalResultSet = finalResultSet
// .Where(item => item.MatchingPoints > 0)
// .GroupBy(item => item.Category)
// .Select(item =>
// new MatcherResult()
// {
// Category = item.Key,
// MatchingPoints = item.Sum(result => result.MatchingPoints),
// MatchDistance = item.Sum(result => result.MatchDistance)
// })
// .ToList();
//}
double totalWeightDistance = finalResultSet.Sum(item => item.MatchDistanceWeight);
double totalWeightPoints = finalResultSet.Sum(item => item.MatchingPointsWeight);
//finalResultSet = finalResultSet
// .Where(item => item.MatchingPoints > 0)
// .GroupBy(item => item.Category)
// .Select(item =>
// new MatcherResult()
// {
// Category = item.Key,
// MatchingPoints = item.Sum(result => result.MatchingPoints * result.MatchingPointsWeight) / totalWeightPoints,
// MatchDistance = item.Sum(result => result.MatchDistance * result.MatchDistanceWeight) / totalWeightDistance
// })
// .ToList();
finalResultSet = finalResultSet
.Where(item => item.MatchingPoints > 0)
.GroupBy(item => item.Category)
.Select(item =>
new MatcherResult()
{
Category = item.Key,
MatchingPoints = item.Sum(result => result.MatchingPointsWeight) / totalWeightPoints,
MatchDistance = item.Sum(result => result.MatchDistanceWeight) / totalWeightDistance
})
.ToList();
finalResultSet = finalResultSet.Where(item => item.MatchingPoints > 0).OrderByDescending(item => item, new MatcherResultWeightedComparer()).Take(3).ToList();
return(finalResultSet);
}
}
private List <MDMatch> clusterBasedOnPoseEstimation(List <MDMatch> matches, KeyPoints queryKeyPoints, PreProcessedImage trainingData)
{
// If no training data is provided then we cannot compute pose (LeafAnalysisV1), hence just return back the original set
if (trainingData == null ||
matches == null ||
!matches.Any())
{
return(matches);
}
Dictionary <MDMatch, List <MDMatch> > clusters = new Dictionary <MDMatch, List <MDMatch> >(matches.Count);
List <PoseEstimate> poseEstimates = new List <PoseEstimate>(matches.Count);
foreach (MDMatch match in matches)
{
MKeyPoint queryKeyPoint = queryKeyPoints.Points[match.QueryIdx];
MKeyPoint trainingKeyPoint = trainingData.KeyPoints.Points[match.TrainIdx];
PoseEstimate estimate = new PoseEstimate();
estimate.Match = match;
estimate.Dx = trainingKeyPoint.Point.X - queryKeyPoint.Point.X;
estimate.Dy = trainingKeyPoint.Point.Y - queryKeyPoint.Point.Y;
estimate.Ds = trainingKeyPoint.Octave / queryKeyPoint.Octave;
estimate.Do = trainingKeyPoint.Angle - queryKeyPoint.Angle;
poseEstimates.Add(estimate);
// Initialize clusters for each individual match
// Next we will add other matches which belong to this cluster
clusters.Add(match, new List <MDMatch>(new MDMatch[] { match }));
}
const double errorThreshold = 5;
// Compute cluster membership
foreach (PoseEstimate estimate in poseEstimates)
{
foreach (PoseEstimate otherEstimate in poseEstimates)
{
// Ignore self
if (estimate == otherEstimate)
{
continue;
}
double error = estimate.RMSE(otherEstimate);
//Console.WriteLine("Error: " + trainingData.Category + ": " + error);
if (error < errorThreshold)
{
clusters[estimate.Match].Add(otherEstimate.Match);
}
}
}
// Finally pick the largest cluster
List <MDMatch> result = null;
int sizeOfCluster = -1;;
foreach (KeyValuePair <MDMatch, List <MDMatch> > cluster in clusters)
{
if (cluster.Value.Count == sizeOfCluster)
{
// Tie breaker: choose the cluster with smaller overall distances
if (result.Sum(item => item.Distance) > cluster.Value.Sum(item => item.Distance))
{
result = cluster.Value;
}
}
else if (cluster.Value.Count > sizeOfCluster)
{
sizeOfCluster = cluster.Value.Count;
result = cluster.Value;
}
}
return(result);
}
public MatcherResult knnMatch(KeyPoints queryDescriptor, BFMatcher matcher, string leafCategory, int distanceCutoff = int.MaxValue, PreProcessedImage trainingData = null)
{
MatcherResult result = new MatcherResult();
result.Category = leafCategory;
using (VectorOfVectorOfDMatch vectorMatchesForSift = new VectorOfVectorOfDMatch())
{
matcher.KnnMatch(queryDescriptor.Descriptor, vectorMatchesForSift, 2, null);
int numberOfMatches = 0;
Dictionary <int, int> counts = new Dictionary <int, int>();
List <MDMatch> goodMatches = new List <MDMatch>(vectorMatchesForSift.Size);
for (int i = 0; i < vectorMatchesForSift.Size; i++)
{
// Do Ratio test: Reject matches where ratio of closest match with second closest if greater than 0.8
if (
(vectorMatchesForSift[i].Size == 1 ||
vectorMatchesForSift[i][0].Distance < 0.75 * vectorMatchesForSift[i][1].Distance) &&
vectorMatchesForSift[i][0].Distance < distanceCutoff)
{
goodMatches.Add(vectorMatchesForSift[i][0]);
numberOfMatches++;
}
}
//goodMatches = clusterBasedOnPoseEstimation(goodMatches, queryDescriptor, trainingData);
int maxResults = int.MaxValue;
result.MatchingPoints = goodMatches.Count;
if (!goodMatches.Any())
{
result.MatchDistance = float.MaxValue;
result.AverageDistance = 0;
result.MatchDistanceWeight = 0;
result.MatchingPointsWeight = 0;
}
else
{
result.MatchDistance = goodMatches.OrderBy(item => item.Distance).Take(maxResults).Sum(item => item.Distance);
result.AverageDistance = result.MatchDistance / result.MatchingPoints;
result.MatchDistanceWeight = 1 / Math.Pow(result.AverageDistance, 2);
result.MatchingPointsWeight = result.MatchingPoints * result.MatchingPoints;
}
}
return(result);
}
