PCA Based Geometric Modeling 1 The 11th International Conference on Computational Science and Its Applications (ICCSA 2011), June 20-23, 2011, Santander, Spain, Accepted PCA Based Geometric Modeling for Automatic Face Detection Authors : Padma Polash Paul Presented by Padma Polash Paul April 13, 2015 Outline 2 Introduction Face Detection Skin Color Model Challenges in Face Detection Background Study Proposed Method Experimental Results Conclusion and Future Work PCA Based Geometric Modeling April 13, 2015 Face Detection vs. Skin Color Model 3 PCA Based Geometric Modeling April 13, 2015 Challenges for Face Detection 4 Face Orientation Multiple view Face Background Time in Massive Processing PCA Based Geometric Modeling April 13, 2015 Color Model 5 Color models for Images RGB, HSV, YCbCr and CIE-Lab etc RGB triple component (RED,GREEN,BLUE) RGB represents not only color but also luminance. Luminance: may vary across a person's face PCA Based Geometric Modeling April 13, 2015 Sink Color Model 6 The common RGB representation of color images is not suitable for characterizing skin-color. Chromatic colors , also known as "pure" colors in the absence of luminance Normalized color for each pixel R, G and B can be define as chromatic color PCA Based Geometric Modeling April 13, 2015 Sink Color Model (Cont’d) 7 Normalized Color Normalized R R (r ) RG B Normalized G G ( g) RG B Normalized ( PCA Based Geometric Modeling b) 1 r g r g b 1 April 13, 2015 Sink Color Model (Cont’d) 8 If two points P1 [r1, g1, b1] and P2 [r2, g2, b2], are proportional, then r1 g1 b1 r2 g 2 b2 Then, P1 and P2 have the same color but different brightness. Chromatic colors are well suited to segment skin regions from non-skin regions. PCA Based Geometric Modeling April 13, 2015 Sink Color Model (Cont’d) 9 Skin color distribution can be represented by Gaussian Model N (m, c) Mean, m E{x} CoVar [ x (rb)T ] rr c, gr rg gg From Gaussian fitted skin color model, we can find the likelihood of skin for any pixel of an image. Establish the threshold for Skin and non skin regions PCA Based Geometric Modeling April 13, 2015 Skin Region Segmentation 10 Algorithm for Skin Region Segmentation using SCM Input image or Video (RGB) Converting into Chromatic Color Space Thresholding image using Skin Color Threshold r = 0.38-0.52 Multiply main RGB Image by Black and white template g = 0.23- 0.34 Apply Region Growing Algorithm PCA Based Geometric Modeling Segmented RGB skin Regions Generate Black and White template for skin regions April 13, 2015 Face Detection 11 PCA Based Geometric Modeling April 13, 2015 Face Detector 12 Proposed Geometric Face Detector PCA Based Geometric Modeling April 13, 2015 Existing Geometric Model 13 Triangle shape Problem Detect non human face as face PCA Based Geometric Modeling April 13, 2015 Geometry of Face 14 Geometric Shape of face Masking More complex interior structure are estimated PCA Based Geometric Modeling April 13, 2015 Geometric Modeling of Face 15 Block diagram of the proposed system Detected Skin Regions Converting into Common Resolution Masking Projecting of PCs Calculate PCA Reconstructing Using Smaller Numbers of PCs Detect Edge Using Canny Edge Detector Estimate Threshold Values For Face and Non Face PCA Based Geometric Modeling Normalized the cumulative sum in the rage [0 1] to get the threshold values for face and non face April 13, 2015 Experimental Result 16 Database Used California Institute of Technology (CIT) Baoface dataset (BaoFace) vision group of Essex University Face Database (Essex), Georgia Tech Face Database (Georgia Tech) PCA Based Geometric Modeling April 13, 2015 Experimental Result 17 Face Detection Accuracy (% of Face) Proposed Method Dataset Accurate Detection Rate (%) Threshold limits Face Detection Accuracy (% of Face) Existing Best Method Accurate Detection Rate (%) CIT 98.7 0.35-0.59 91.2 BaoFace 97.1 0.24-0.39 92.5 Essex 97.1 0.39-0.52 93.7 Georgia Tech 96.7 0.25-0.45 85.2 PCA Based Geometric Modeling April 13, 2015 System Performance Result 18 For video or larger database Processing time for face detection is important Skin region detection is fast because of the thresholding. Time complexity of the Face detection system is O(1). After edge detection cumulative sum is compared Threshold value is rotation invariant because we are taking the cumulative sum of the projected geometric structure PCA Based Geometric Modeling April 13, 2015 Conclusion and Future Work 19 We presented a new method based on modeling of geometric structure of the face method for automatic face detection. Fusion of PCA based geometric modeling and SCM method provides higher face detection accuracy and improves time complexity. In the future, using more complex geometric structure can be used for better understanding of the important facial features and threshold values. Complex structure will also help to obtain a better and more generalized threshold for the face. PCA Based Geometric Modeling April 13, 2015 20 Thanks Any Questions? PCA Based Geometric Modeling April 13, 2015