public void ClassifySegments()
{
//label some segments as noise
//label the color group with the largest segment as a background element
//label all segments in the color group as a background (?), unless it is noise
//001-noise, 010-background, 100-foreground
double largestSize = 0;
int segIdx = -1;
for (int i = 0; i < segments.Count(); i++)
{
double size = segments[i].points.Count();
if (segments[i].assignmentId >= 0 && largestSize < size)
{
largestSize = size;
segIdx = i;
}
}
int groupId = segments[segIdx].groupId;
NamedFeature bg = new NamedFeature("Label");
bg.values = new List <Double> {
0, 1, 0
};
foreach (int idx in groups[groupId].members)
{
if (segments[idx].assignmentId >= 0)
{
segments[idx].features.Add(bg);
}
}
NamedFeature noise = new NamedFeature("Label");
noise.values = new List <Double> {
0, 0, 1
};
NamedFeature fg = new NamedFeature("Label");
fg.values = new List <Double> {
1, 0, 0
};
foreach (Segment s in segments)
{
if (s.assignmentId < 0)
{
s.features.Add(noise);
}
else if (s.groupId != groupId)
{
s.features.Add(fg);
}
}
}
public void ComputeFeatures(SegmentGroup g)
{
double imarea = (double)(imageWidth * imageHeight);
// Amount of the image this group takes up
NamedFeature sizef = new NamedFeature("RelativeSize");
int size = 0;
foreach (int i in g.members)
{
Segment s = segments[i];
size += s.points.Count();
}
sizef.values.Add(size / imarea);
g.features.Add(sizef);
// (Normalized) number of segments in the group
NamedFeature numsegf = new NamedFeature("NormalizedNumSegs");
numsegf.values.Add(g.members.Count() / (double)(segments.Count()));
g.features.Add(numsegf);
// min, max, mean, stddev segment size
int minsegsize, maxsegsize, segsizesum;
minsegsize = Int32.MaxValue;
maxsegsize = 0;
segsizesum = 0;
foreach (int i in g.members)
{
int s = segments[i].points.Count;
minsegsize = Math.Min(s, minsegsize);
maxsegsize = Math.Max(s, maxsegsize);
segsizesum += s;
}
double meansegsize = ((double)segsizesum) / g.members.Count;
double stddevsegsize = 0;
foreach (int i in g.members)
{
int s = segments[i].points.Count;
double diff = (s - meansegsize);
stddevsegsize += (diff * diff);
}
stddevsegsize /= g.members.Count;
NamedFeature segsizestatsf = new NamedFeature("SegmentSizeStatistics");
segsizestatsf.values.Add(minsegsize / imarea);
segsizestatsf.values.Add(maxsegsize / imarea);
segsizestatsf.values.Add(meansegsize / imarea);
segsizestatsf.values.Add(stddevsegsize / imarea);
g.features.Add(segsizestatsf);
// "spread" - covariance of segment centroids
PointF meanc = new PointF(0, 0);
foreach (int i in g.members)
{
PointF c = NormalizedCentroid(segments[i]);
meanc.X += c.X;
meanc.Y += c.Y;
}
meanc.X /= g.members.Count;
meanc.Y /= g.members.Count;
double[,] covar = new double[, ] {
{ 0, 0 }, { 0, 0 }
};
foreach (int i in g.members)
{
PointF c = NormalizedCentroid(segments[i]);
float x = c.X - meanc.X;
float y = c.Y - meanc.Y;
covar[0, 0] += x * x;
covar[0, 1] += x * y;
covar[1, 0] += x * y;
covar[1, 1] += y * y;
}
covar[0, 0] /= g.members.Count;
covar[0, 1] /= g.members.Count;
covar[1, 0] /= g.members.Count;
covar[1, 1] /= g.members.Count;
NamedFeature spreadf = new NamedFeature("SegmentSpread",
new List <double> {
covar[0, 0], covar[0, 1], covar[1, 0], covar[1, 1]
});
g.features.Add(spreadf);
}
public void ComputeFeatures(Segment s)
{
//Add the relative size
NamedFeature f = new NamedFeature("RelativeSize");
f.values.Add(s.points.Count() / (double)(imageWidth * imageHeight));
s.features.Add(f);
// Relative centroid
PointF c = NormalizedCentroid(s);
s.features.Add(new NamedFeature("RelativeCentroid", new List <double> {
c.X, c.Y
}));
// One interior point
PointF np = RandomNormalizedInteriorPoint(s);
s.features.Add(new NamedFeature("OneInteriorPoint", new List <double> {
np.X, np.Y
}));
//Radial distance
s.features.Add(new NamedFeature("RadialDistance", new List <double> {
Math.Sqrt(c.X * c.X + c.Y * c.Y)
}));
//Normalized Discrete Compactness http://www.m-hikari.com/imf-password2009/25-28-2009/bribiescaIMF25-28-2009.pdf
//Find the segment id
Point sp = s.points.First();
int sidx = assignments[sp.X, sp.Y];
//count number of perimeter edges
int perimeter = 0;
foreach (Point p in s.points)
{
for (int i = -1; i <= 1; i++)
{
for (int j = -1; j <= 1; j++)
{
if (Math.Abs(i) == Math.Abs(j))
{
continue;
}
if (Util.InBounds(p.X + i, p.Y + j, imageWidth, imageHeight) && assignments[p.X + i, p.Y + j] != sidx)
{
perimeter++;
}
else if (!Util.InBounds(p.X + i, p.Y + j, imageWidth, imageHeight)) //edge pixels should be considered perimeter too
{
perimeter++;
}
}
}
}
int n = s.points.Count();
double CD = (4.0 * n - perimeter) / 2;
double CDmin = n - 1;
double CDmax = (4 * n - 4 * Math.Sqrt(n)) / 2;
double CDN = (CD - CDmin) / Math.Max(1, (CDmax - CDmin));
s.features.Add(new NamedFeature("NormalizedDiscreteCompactness", new List <double> {
CDN
}));
//Add elongation (width/length normalized between 0-square to 1-long http://hal.archives-ouvertes.fr/docs/00/44/60/37/PDF/ARS-Journal-SurveyPatternRecognition.pdf
PointF[] points = s.points.Select <Point, PointF>(p => new PointF(p.X, p.Y)).ToArray <PointF>();
Emgu.CV.Structure.MCvBox2D box = Emgu.CV.PointCollection.MinAreaRect(points);
PointF[] vertices = box.GetVertices();
double elongation = 1 - Math.Min(box.size.Width + 1, box.size.Height + 1) / Math.Max(box.size.Width + 1, box.size.Height + 1);
s.features.Add(new NamedFeature("Elongation", new List <double> {
elongation
}));
//Add Hu shape moments, invariant to translation, scale, and rotation (not sure what each measure refers to intuitively though, or if there is an intuitive analog)
//They may also do badly on noisy data however. See: http://hal.archives-ouvertes.fr/docs/00/44/60/37/PDF/ARS-Journal-SurveyPatternRecognition.pdf (called Invariant Moments)
Bitmap regionBitmap = new Bitmap(imageWidth, imageHeight);
Graphics g = Graphics.FromImage(regionBitmap);
g.FillRectangle(new SolidBrush(Color.Black), 0, 0, imageWidth, imageHeight);
foreach (Point p in s.points)
{
regionBitmap.SetPixel(p.X, p.Y, Color.White);
}
Emgu.CV.Image <Gray, byte> region = new Emgu.CV.Image <Gray, byte>(regionBitmap);
MCvMoments moment = region.GetMoments(true);
MCvHuMoments hu = moment.GetHuMoment();
s.features.Add(new NamedFeature("HuMoments", new List <double> {
hu.hu1, hu.hu2, hu.hu3, hu.hu4, hu.hu5, hu.hu6, hu.hu7
}));
region.Dispose();
regionBitmap.Dispose();
}
public void ComputeFeatures(SegmentGroup g)
{
double imarea = (double)(imageWidth * imageHeight);
// Amount of the image this group takes up
NamedFeature sizef = new NamedFeature("RelativeSize");
int size = 0;
foreach (int i in g.members)
{
Segment s = segments[i];
size += s.points.Count();
}
sizef.values.Add(size / imarea);
g.features.Add(sizef);
// (Normalized) number of segments in the group
NamedFeature numsegf = new NamedFeature("NormalizedNumSegs");
numsegf.values.Add(g.members.Count() / (double)(segments.Count()));
g.features.Add(numsegf);
// min, max, mean, stddev segment size
int minsegsize, maxsegsize, segsizesum;
minsegsize = Int32.MaxValue;
maxsegsize = 0;
segsizesum = 0;
foreach (int i in g.members)
{
int s = segments[i].points.Count;
minsegsize = Math.Min(s, minsegsize);
maxsegsize = Math.Max(s, maxsegsize);
segsizesum += s;
}
double meansegsize = ((double)segsizesum) / g.members.Count;
double stddevsegsize = 0;
foreach (int i in g.members)
{
int s = segments[i].points.Count;
double diff = (s - meansegsize);
stddevsegsize += (diff * diff);
}
stddevsegsize /= g.members.Count;
NamedFeature segsizestatsf = new NamedFeature("SegmentSizeStatistics");
segsizestatsf.values.Add(minsegsize / imarea);
segsizestatsf.values.Add(maxsegsize / imarea);
segsizestatsf.values.Add(meansegsize / imarea);
segsizestatsf.values.Add(stddevsegsize / imarea);
g.features.Add(segsizestatsf);
// "spread" - covariance of segment centroids
PointF meanc = new PointF(0, 0);
foreach (int i in g.members)
{
PointF c = NormalizedCentroid(segments[i]);
meanc.X += c.X;
meanc.Y += c.Y;
}
meanc.X /= g.members.Count;
meanc.Y /= g.members.Count;
double[,] covar = new double[,] { { 0, 0 }, { 0, 0 } };
foreach (int i in g.members)
{
PointF c = NormalizedCentroid(segments[i]);
float x = c.X - meanc.X;
float y = c.Y - meanc.Y;
covar[0, 0] += x * x;
covar[0, 1] += x * y;
covar[1, 0] += x * y;
covar[1, 1] += y * y;
}
covar[0, 0] /= g.members.Count;
covar[0, 1] /= g.members.Count;
covar[1, 0] /= g.members.Count;
covar[1, 1] /= g.members.Count;
NamedFeature spreadf = new NamedFeature("SegmentSpread",
new List<double>{covar[0,0], covar[0,1], covar[1,0], covar[1,1]});
g.features.Add(spreadf);
}
public void ComputeFeatures(Segment s)
{
//Add the relative size
NamedFeature f = new NamedFeature("RelativeSize");
f.values.Add(s.points.Count() / (double)(imageWidth * imageHeight));
s.features.Add(f);
// Relative centroid
PointF c = NormalizedCentroid(s);
s.features.Add(new NamedFeature("RelativeCentroid", new List<double>{c.X, c.Y}));
// One interior point
PointF np = RandomNormalizedInteriorPoint(s);
s.features.Add(new NamedFeature("OneInteriorPoint", new List<double> { np.X, np.Y }));
//Radial distance
s.features.Add(new NamedFeature("RadialDistance", new List<double>{Math.Sqrt(c.X*c.X+c.Y*c.Y)}));
//Normalized Discrete Compactness http://www.m-hikari.com/imf-password2009/25-28-2009/bribiescaIMF25-28-2009.pdf
//Find the segment id
Point sp = s.points.First();
int sidx = assignments[sp.X, sp.Y];
//count number of perimeter edges
int perimeter = 0;
foreach (Point p in s.points)
{
for (int i = -1; i <= 1; i++)
{
for (int j = -1; j <= 1; j++)
{
if (Math.Abs(i) == Math.Abs(j))
continue;
if (Util.InBounds(p.X + i, p.Y + j, imageWidth, imageHeight) && assignments[p.X + i, p.Y + j] != sidx)
perimeter++;
else if (!Util.InBounds(p.X + i, p.Y + j, imageWidth, imageHeight)) //edge pixels should be considered perimeter too
perimeter++;
}
}
}
int n = s.points.Count();
double CD = (4.0 * n - perimeter) / 2;
double CDmin = n - 1;
double CDmax = (4 * n - 4 * Math.Sqrt(n)) / 2;
double CDN = (CD - CDmin) / Math.Max(1,(CDmax - CDmin));
s.features.Add(new NamedFeature("NormalizedDiscreteCompactness", new List<double> { CDN }));
//Add elongation (width/length normalized between 0-square to 1-long http://hal.archives-ouvertes.fr/docs/00/44/60/37/PDF/ARS-Journal-SurveyPatternRecognition.pdf
PointF[] points = s.points.Select<Point, PointF>(p => new PointF(p.X, p.Y)).ToArray<PointF>();
Emgu.CV.Structure.MCvBox2D box = Emgu.CV.PointCollection.MinAreaRect(points);
PointF[] vertices = box.GetVertices();
double elongation = 1 - Math.Min(box.size.Width + 1, box.size.Height + 1) / Math.Max(box.size.Width + 1, box.size.Height + 1);
s.features.Add(new NamedFeature("Elongation", new List<double>{elongation}));
//Add Hu shape moments, invariant to translation, scale, and rotation (not sure what each measure refers to intuitively though, or if there is an intuitive analog)
//They may also do badly on noisy data however. See: http://hal.archives-ouvertes.fr/docs/00/44/60/37/PDF/ARS-Journal-SurveyPatternRecognition.pdf (called Invariant Moments)
Bitmap regionBitmap = new Bitmap(imageWidth, imageHeight);
Graphics g = Graphics.FromImage(regionBitmap);
g.FillRectangle(new SolidBrush(Color.Black), 0, 0, imageWidth, imageHeight);
foreach (Point p in s.points)
{
regionBitmap.SetPixel(p.X, p.Y, Color.White);
}
Emgu.CV.Image<Gray, byte> region = new Emgu.CV.Image<Gray, byte>(regionBitmap);
MCvMoments moment = region.GetMoments(true);
MCvHuMoments hu = moment.GetHuMoment();
s.features.Add(new NamedFeature("HuMoments", new List<double> {hu.hu1, hu.hu2, hu.hu3,hu.hu4,hu.hu5, hu.hu6, hu.hu7 }));
region.Dispose();
regionBitmap.Dispose();
}
public void ClassifySegments()
{
//label some segments as noise
//label the color group with the largest segment as a background element
//label all segments in the color group as a background (?), unless it is noise
//001-noise, 010-background, 100-foreground
double largestSize = 0;
int segIdx = -1;
for (int i = 0; i < segments.Count(); i++)
{
double size = segments[i].points.Count();
if (segments[i].assignmentId >= 0 && largestSize < size)
{
largestSize = size;
segIdx = i;
}
}
int groupId = segments[segIdx].groupId;
NamedFeature bg = new NamedFeature("Label");
bg.values = new List<Double>{0, 1, 0};
foreach (int idx in groups[groupId].members)
{
if (segments[idx].assignmentId >=0)
segments[idx].features.Add(bg);
}
NamedFeature noise = new NamedFeature("Label");
noise.values = new List<Double>{0, 0, 1};
NamedFeature fg = new NamedFeature("Label");
fg.values = new List<Double> { 1, 0, 0 };
foreach (Segment s in segments)
{
if (s.assignmentId < 0)
s.features.Add(noise);
else if (s.groupId != groupId)
s.features.Add(fg);
}
}
