Slide 1Slide 2Slide 3Slide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Slide 13Slide 14Face Detection In Color ImagesWenmiao Lu Shaohua SunGroup 3EE368 Project•Overview•Human Skin Segmentation•Adaptive Shape Analysis•View-based Face Detection•ResultsEE368 ProjectSkin Segmentation Shape Analysis Face DetectionHuman Skin SegmentationEE368 Project•Use YCbCr color space for good cluster separation•Model the skin and background color distributions with GMM•Segmentation by maximum likelihood classificationAn Example for Initial Skin SegmentationEE368 ProjectFairly complete skin segmentation with some noiseAdaptive Shape AnalysisEE368 ProjectRefine the binary map Open to get smaller regions Initial Face IdentificationErosion & Dilation Different Structuring ElementsPrior InformationAn Example for Adaptive Shape AnalysisEE368 Project•Medium size: faces•Small, big or odd shaped regions: passed to next stageView-Based Face DetectionEE368 ProjectProject to Low-dimensional Feature Space Spanned by Largest EigenvectorsFace/Non-Face DecisionTest PatternDistances to Face ModelEE368 ProjectTest pattern is measured against the Face Model, which consists ofi) 6 Face Clusters and ii) 6 Non-face Clusters*Figure is obtain from Sung, Kah Kay (1996)Learning and Example Selection for Object and Pattern Detection.Ph.D. Thesis, Massachusetts Institute of Technology, 1995.Distances between Test Pattern and One ClusterEE368 Project*Figure is obtain from Sung, Kah Kay (1996)Learning and Example Selection for Object and Pattern Detection.Ph.D. Thesis, Massachusetts Institute of Technology, 1995.Neural Network ClassificationEE368 Project• 2-distance metric is discriminative for face and non-face patterns.• 2 distances have different magnitude; neural network performs the final classification.*Figure is obtain from Sung, Kah Kay (1996)Learning and Example Selection for Object and Pattern Detection.Ph.D. Thesis, Massachusetts Institute of Technology, 1995.Experimental ResultsDetection Rate: 95.6% False Positive: 0.6%EE368 ProjectEE368
View Full Document