EE368 Face detectionIntroductionColor SegmentationSlide 4Image SegmentationRoberts cross edge detectionSlide 7Image MatchingImage Matching (cont.)Slide 10ResultsConclusionSlide 13Slide 14Stanford UniversityEE368 Face detectionJoon Hyung Shim, Jinkyu Yang, and Inseong Kim2IntroductionFace detection : important part of face recognitionVariations of image appearance pose (front, non-front)occlusionImage orientationilluminating conditionfacial expression. Methods color segmentation image segmentationtemplate matching methods.3Color Segmentation RGB components YCbCr components Y = 0.299R + 0.587G + 0.114B Cb = -0.169R - 0.332G + 0.500B Cr = 0.500R - 0.419G - 0.081B Skin window : from mean, deviation of Cb, Cr components.Skin pixel of YCbCr color space4Color segmentation result of a training image5Image SegmentationSeparate the image blobs into individual regionsFill up black isolated holes, remove white isolated regionSeparate some integrated regions into individual facesRoberts cross edge detection algorithmHighlights regions(edge) black line erodePrevious images are integrated into one binary image Small black and white areas are removed.6Roberts cross edge detection gradient magnitude : |G | = ( Gx2 + Gy2 ) ½ or |G | = |Gx | + |Gy | Angle of orientation : θ = arctan (Gy /Gx ) - 3π/4 Pseudo-convolution operator magnitude : |G | = |P1 – P4 | + |P2 – P3 |Pseudo-Convolution masks Roberts Cross convolution masks7 Preliminary face detection with red marks8Image MatchingEigenimage Generation10 eigenimages using 106 test Average image using eigenimagesBuilding Eigenimage Database30 220 pixel-width square image with 10-pixel gap9Image Matching (cont.)Test Image Selection : box-merge algorithm Merging of Adjacent Boxes Correlation : image matching algorithm Normalized test image : gray , average brightness of skin colorDistance compensation->10Image Matching (cont.)Filtering using Statistical Information : non-face removal Histogram : Imaging matching Correlation Ranking after Geographical Consideration11Results Face Detection Results using 7 Training Images Right hit rate : 93.3 % Repeat rate : 0 %False hit rate : 4.2 % The average run time : 96 seconds.numFaces numHitnumRepeatnumFalserun time[sec]Training_1.jpg 21 19 0 0 111Training_2.jpg 24 24 0 1 101Training_3.jpg 25 23 0 1 89Training_4.jpg 24 21 0 1 84Training_5.jpg 24 22 0 0 93Training_6.jpg 24 22 0 3 100Training_7.jpg 22 22 0 1 9512ConclusionColor segmentation Rectangular window must be in actual distribution of skin color Image segmentation Unnecessary noises in edge integration Roberts cross operator : small-hole removalSobel cross filter, prewitt filter Threshold : discriminate face edges from other edge lines effectivelySkin-colored areas : unnecessary squares one faceEigenimage matchingStatistical approachSophisticated algorithm for general
View Full Document