Color Segmentation

1. Color 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 image

2. Hue and Saturation Faces Training Image Q Distribution

3. Mask After Color Segmentation • Skin elements remain. • Holes in faces later eliminated with hole-filling

4. Mask After Object Removal Based on size distribution of remaining objects, remove small ones

5. Correlation 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.

6. Correlation 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 zero

7. Average face template Poor separation between faces Difficult to identify face centroid Edge face template Better separation between faces Peaks (centroid) more easily identifiable Correlation comparison

8. Region counting - Supplementary method • The edge outlines have clearly identifiable connected regions • Can be counted, and statistics used to help reject clutter Number of regions: 14 Number of regions: 43

9. Correlation Dimensions Region counting Detection Algorithm • Correlation – Degree of matching • Dimensions – height, width • Region counting – complexity of image Single face Multiple faces Correlation Dimensions Region counting Multi-face detection

10. Multiple 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 repeats

11. Find 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 faces

12. Next 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 found

13. Find 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 location

14. Conclusion • 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 estimate)