Face Detection on Similar Color ImagesProblem StatementSample ImageDetection ProcedureSkin DetectionSkin Detection (cont’d)Histograms2 Dimensional PDF3 Dimensional PDF“Closing” StepTemplate MatchingTrial TemplatesFace SelectionResultsConclusionsSlide 16Slide 17Slide 18EE368: Digital Image ProcessingBernd GirodLeahy, p.1/15Face Detection on Similar Color ImagesScott LeahyEE368, Stanford UniversityMay 30, 2003EE368: Digital Image ProcessingBernd GirodLeahy, p.2/15Problem Statement•Goal: Find faces in an image–All images are in color–Images all contain a similar background–Images have a similar number of faces–Faces are all on approximately the same scale•Design an algorithm which takes advantage of these factsEE368: Digital Image ProcessingBernd GirodLeahy, p.3/15Sample ImageEE368: Digital Image ProcessingBernd GirodLeahy, p.4/15Detection Procedure•Steps Involved:–Skin Detection–Morphological Processing–Template Matching–Face Coordinate SelectionEE368: Digital Image ProcessingBernd GirodLeahy, p.5/15Skin Detection•Pixel by pixel, make a decision on the input based on the output–i = {skin, non-skin}–v = vector in color space (HSV, RGB, …)•Treat the problem like a digital communications problem–Create a MAP Detector?i vEE368: Digital Image ProcessingBernd GirodLeahy, p.6/15Skin Detection (cont’d)•MAP Detection–Minimize probability of error:•Maximize p(i|v) over all inputs i–Often p(i|v) is not known, but:•p(i|v) = p(v|i) * p(i) / p(v) (Bayes’ Rule)–p(v|i) and p(i) are more often known in a systemEE368: Digital Image ProcessingBernd GirodLeahy, p.7/15HistogramsEE368: Digital Image ProcessingBernd GirodLeahy, p.8/152 Dimensional PDF•Used only Hue and Saturation for MAP detectorEE368: Digital Image ProcessingBernd GirodLeahy, p.9/153 Dimensional PDF•Used all 3 coordinates for MAP detectorEE368: Digital Image ProcessingBernd GirodLeahy, p.10/15“Closing” Step•Pseudo-Closing Step:–Dilation–Filling–ErosionEE368: Digital Image ProcessingBernd GirodLeahy, p.11/15Template Matching•Template matching involves convolving the image with some template–The average of the image being tested must be subtracted to eliminate biasing toward brighter areas•Only one template used due to similar size and shape of faces in all imagesEE368: Digital Image ProcessingBernd GirodLeahy, p.12/15Trial Templates•Tried 4 templates, tweaking threshold until the best results were obtainedEE368: Digital Image ProcessingBernd GirodLeahy, p.13/15Face Selection•Labeled all regions•Selected only regions with areas bigger than some threshold•Found the centers of the remaining regions and returned those as the results of the algorithmEE368: Digital Image ProcessingBernd GirodLeahy, p.14/15ResultsEE368: Digital Image ProcessingBernd GirodLeahy, p.15/15Conclusions•Skin Detection and Closing–Takes advantage of images being in color–Takes advantage of similar statistics in the images•Template Matching and Face Selection–Takes advantage of similar size and shape to faces•Result: ~85% success rateEE368: Digital Image ProcessingBernd GirodLeahy, p.16/15EE368: Digital Image ProcessingBernd GirodLeahy, p.17/15EE368: Digital Image ProcessingBernd GirodLeahy,
View Full Document