private static List<DistrPair> prepareDistribution(Segment[] segments, Photo photo) {
List<DistrPair> distribution = new ArrayList<>();
int area = photo.getWidth() * photo.getHeight();
double sum = 0;
for (Segment seg : segments) {
DistrPair dp = new DistrPair();
// dp.weight = (double)seg.getSumPixels() / area;
/* Original method uses sqrt */
dp.weight = Math.sqrt((double) seg.getSumPixels() / area);
sum += dp.weight;
Vector vect = new Vector(14);
Addterator<Float> addter = vect.addterator();
/* Color moments - 9 elements in total */
float[] mean_hsb = seg.getMean();
float[] stdev_hsb = seg.getStdDeviation();
float[] skew_hsb = seg.getSkewness();
for (int i = 0; i < Segment.MODEL_NUM_ELEMENTS; i++) {
addter.add(mean_hsb[i]);
addter.add(stdev_hsb[i]);
addter.add(skew_hsb[i]);
}
/* The other elements - another 5 of them */
int o_width = seg.getRight() - seg.getLeft();
o_width = Math.max(o_width, 1);
int o_height = seg.getBottom() - seg.getTop();
o_height = Math.max(o_height, 1);
int o_area = o_width * o_height;
/* 5 elements characterizing the shape of one segment */
addter.add((float) Math.log((double) o_width / o_height + 1));
addter.add((float) Math.log((double) o_area / area + 1));
addter.add((float) seg.getSumPixels() / o_area);
addter.add((float) seg.getX() / photo.getWidth());
addter.add((float) seg.getY() / photo.getHeight());
dp.vector = vect;
distribution.add(dp);
}
/* Scale the sum of weights to be 1 */
for (DistrPair distrPair : distribution) {
distrPair.weight /= sum;
}
return distribution;
}
/**
* For given two photos, count their similarity. The result is from interval [0;1], 1 means
* perfect match, 0 totaly different.
*
* @param photo1 First photo.
* @param photo2 Second photo.
* @return Similarity of given photos from interval [0;1];
* @throws com.kappa_labs.ohunter.lib.net.OHException When Photos are wrongly set.
*/
public static synchronized float computeSimilarity(Photo photo1, Photo photo2)
throws OHException {
float ret = 0;
/* Scale the images to provide the best results */
try {
photo1.sImage = photo1._sImage = new SImage(photo1.sImage);
photo2.sImage = photo2._sImage = new SImage(photo2.sImage);
((SImage) photo1.sImage).setImage(resize(((SImage) photo1._sImage).toBufferedImage()));
((SImage) photo2.sImage).setImage(resize(((SImage) photo2._sImage).toBufferedImage()));
} catch (Exception e) {
LOGGER.log(Level.WARNING, "Could not acquire photos from client: {0}", e);
throw new OHException("Could not acquire photos!", OHException.EXType.OTHER);
}
/* Count average from few attempts */
final int numRepeats = settingsManager.getSimilarityNumberOfRepeats();
int iteration = numRepeats;
while (iteration-- > 0) {
/* Perform segmentation */
Segment[] segments1 = Segmenter.findSegments(photo1);
Segment[] segments2 = Segmenter.findSegments(photo2);
LOGGER.log(
Level.FINER,
"... got {0} segments from first and {1} segments" + " in the second photo",
new Object[] {segments1.length, segments2.length});
/* Create new Problem from given counted segments */
Problem problem = new Problem();
problem.distr1 = prepareDistribution(segments1, photo1);
problem.distr2 = prepareDistribution(segments2, photo2);
/* Solve the EMP linear problem and return the final result */
EMDSolver empm = new EMDSolver(problem);
float act = Math.max(0f, Math.min(1f, (float) empm.countValue()));
LOGGER.finer(String.format(" - similarity: %.1f%%", 100 - act * 100));
ret += act;
}
ret /= numRepeats;
return 1f - ret;
}