Color SegmentationHue and SaturationMask After Color SegmentationMask After Object RemovalCorrelation Template Matching I – Average FaceCorrelation Template Matching II – Edge detectionCorrelation comparisonRegion counting - Supplementary methodDetection AlgorithmMultiple Faces within a Single RegionFind Largest PeakNext PeakFind Multiple FacesConclusionSlide 15Slide 16Color Segmentation•View the YIQ color space:-Y=luminance, I=hue, Q=saturation•Human skin occupy a small portion of the I and Q spaces.•From training images, compare and contrast hue and saturation of:faces only vs. entire imageHue and Saturation-150 -100 -50 0 50 100 15002468101214x 105Histogram of Q Components of Training7.jpg Q DistributionTraining ImageFaces• Skin elements remain.• Holes in faces later eliminated with hole-fillingMask After Color SegmentationMask After Object RemovalBased on size distribution of remaining objects, remove small onesCorrelation Template Matching I – Average Face•First attempt – Average face•Taking average of all faces from ground truth masks•Results – Less than satisfactory. –Face with distinguishing features blurred–Correlation separation is not high, identifies many skin color regions (clothing, background) as false positives. NiixNH11Correlation Template Matching II – Edge detection•After color segmentation, most remaining regions are composed of skin-color tones.•Distinguishing features resides in edges–Use Canny edge filter on black-white images for extraction–Composed average face using edges, scaled to mean zeroCorrelation comparison•Average face template–Poor separation between faces–Difficult to identify face centroid•Edge face template–Better separation between faces–Peaks (centroid) more easily identifiableRegion counting - Supplementary method•The edge outlines have clearly identifiable connected regions•Can be counted, and statistics used to help reject clutterNumber of regions: 14Number of regions: 43Detection Algorithm–Correlation – Degree of matching–Dimensions – height, width–Region counting – complexity of imageCorrelation Dimensions Region countingCorrelation Dimensions Region countingMulti-face detectionSingle faceMultiple facesMultiple Faces within a Single Region•Search for peaks in correlation•A single face may give multiple peaks•Estimate expected number of faces within Region•Do not want repeatsFind Largest Peak•Find largest peak in correlation•Location of first peak•Exclude area of radius R (about peak) from rest of search•R determined dynamically from size of region and number of expected facesNext Peak•Find next largest peak•Exclude area (of radius R) surrounding both peaks from further search•Continue search in this manner until desired number of peaks foundFind Multiple Faces•Stop search if there are no more peaks to be found(Number of peaks found can be fewer than estimate)•Each peak location corresponds to face center locationConclusion•Reasonably successful performance–Misses–False positives/repeats•Algorithm relies heavily on Color Segmentation and Edge Extraction•Difficulty with closely-spaced faces–Separation–Detecting multiple faces in single region (correct
View Full Document