1 / 17

A new face detection method based on shape information

A new face detection method based on shape information. Pattern Recognition Letters, 21 (2000) 463-471. Speaker: M.Q. Jing. Outline. Introduction The overview of the system Image Preprocessing Edge linking Template matching Experiments & Conclusion. Input Image. Contour output.

Download Presentation

A new face detection method based on shape information

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. A new face detection method based on shape information Pattern Recognition Letters, 21 (2000) 463-471 Speaker: M.Q. Jing

  2. Outline • Introduction • The overview of the system • Image Preprocessing • Edge linking • Template matching • Experiments & Conclusion

  3. Input Image Contour output Median filtering Edge detection Edge linking Template matching Image enhancement The overview of the system

  4. Image Preprocessing • Histogram equalization • Median filtering After linking

  5. Tips: 1. The physics meaning 2. Curve1 and Curve2 3. The property of H Edge linking(1/2) -Energy function The energy function of a edge can be defined as following: Curve 1 Curve 2

  6. Edge linking(2/2)-grouping the segments 1. Form the edge chains from the edge image 2. Erase the chain whose the length < ath 3. Decomposing each line contour into straight lines and arcs. 4. For each segment, we search the other segment in a region to make their linkage have optimal H value.

  7. Template matching • In this paper, we use an elliptical ring as the template. • The problem is reduced to finding a elliptic object Included all contour edge point

  8. y r1=unit vector of the edge point r2=unit vector on the ellipse α x Template matching • Define a function: If the contour more like a ellipse, the value is more large. Score=r1 *r2=cos(α)

  9. y r1=unit vector of the edge point r2=unit vector on the ellipse α x O = center of the elliptical A = random point AB= tangent intersecting the positive X-axis at B. • How to find the r1 & r2 vector? r1=the unit vector of tangent of edges at A. = Solbel operator( edge point)

  10. The matching algo. 1.For each set of temple parameters(x0,y0,a,C1) 2. If edge points # in the template < Nt goto 1 3. If the [Score] < Rt goto 1 • Store the [score] with the parameters 5. The parameter with largest [score] is the final result.

  11. Experimental results • Training set: 20 images to get the thresholds • Testing set: MIT face data base: 1. 16 people is digitized 27 times 2. Varying the head orientation, 3. lighting, and the scale 4. 128x120 with complicated background

  12. Detection result on MIT database Correct detection False detection Inexact Exact Our system 50/432 84.96 15/432 Inexact: face is mainly included but non-face region is include or some feature is missed.

  13. Testing set: • The CMU face database, etc… • Total image = 50 (simple background) • Correction location=50 (correction detection=100%) • False detection=8 (Inexact detection rate=16%)

  14. Testing set: • The CMU face database, etc… • Total image = 40 (complex background) • Correction location=35 (correction detection=87.5%) • Can’t find the face=5 (false rate=12.5%) • False detection=5 (Inexact detection rate=12.5%)

  15. Comparison between Govindaraju,1996 and out system Performance system For faces with Shadow, tilting or Bad contrast Correct detection rate Govindaraju,1996 83% Good Our system 87.5% Better • Multi-face detection • At first, Giving the number of face in the image • Select a fixed number of locations with high score as face region

  16. Conclusions - The method can detect the face with simple and complex background - more robust to noise and shape variations. - The template does not include enough info. to distinguish faces in very complex backgrounds.

  17. What are we learning after reading this paper? • A new method for edges connecting can be used in image preprocessing. (ex. Finger printer) • Finding the ellipses in the image. (useful in face detection) • A new note is that the histogram equalization is not only enhancing the image contrast but also the noises.

More Related