1 / 14

D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002

D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002. Face Detection using Fisher Linear Discriminant with A Cross Trained Neural Network (A Multi-Algorithm Approach) Dileep George & Shantanu Rane [ EE 368 Class Project, Spring 2002 ].

justin
Download Presentation

D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Face Detection using Fisher Linear Discriminant with A Cross Trained Neural Network (A Multi-Algorithm Approach) Dileep George & Shantanu Rane [EE 368 Class Project, Spring 2002]

  2. The Two Category Classification Problem Non Ideal Classifier Ideal Classifier N-D Space D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002

  3. ALGO 1 ALGO 2 ALGO 3 D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 The Multi-algorithm approach • Several classification algorithms in pipeline • Different algorithms map N-D space differently to the real line • Thresholds set such that no faces are rejected.

  4. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Multi-algorithm Advantages • Shrinks search space for subsequent algorithms • Reduced computational complexity • Algorithm Parameters can be tuned to the False Alarm Space of its predecessor.

  5. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Algorithm 1: Skin Color Segmentation • YUV Space : Mean Cb, Mean Cr, Cov CbCr • Thresholding mechanism • Number of Operations =~ 1 per pixel. • Followed by Morphological Operations Search-space reduced by a factor of 0.5 [Chang, Robles, EE368 Class Project, Spr 2000]

  6. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Algorithm 2: Fischer Linear Discriminant • Finds best direction of projection w for maximal separation. • 1 convolution per pixel [Duda, Hart, Stork, 2002]

  7. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Search Space After FLD Search-space reduced by a factor of 0.0062

  8. Mahalanobis Distance • Approximated as sum of two components • DIFS • DFFS • Eigenfaces • 8 convolutions per pixel [Moghaddam, Pentland,1997] D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Algorithm 3: Mahalanobis Distance

  9. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Search Space After Mahalanobis Distance Operation Search-space reduced by a factor of 0.0037

  10. 700 400 D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Algorithm 4: Neural Network • Non-linear technique • High complexity - 300 convolutions per pixel. • Gradient Descent with Momentum • Boot-strapping • Cross Algorithm Training

  11. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Reduced Computational Complexity Sample Calculations for a 400-700-1 Neural Network

  12. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Special Cases • Texture Detection • Multi-scale Operation • Hard-to-Detect Faces • False Alarm Reduction

  13. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Results

  14. D.George,S.Rane : Stanford University EE 368 Project Presentation, May 28, 2002 Conclusion • Scale Sensitivity • Representation Vs Discrimination • Cross Algorithm Training • Size of Training Set

More Related