DOOR ENTRY SYSTEM Alina Dinca László Papp Adrian Ulges Csaba Domokos Cercel Constantin Team:

11
DOOR ENTRY SYSTEM Alina Dinca László Papp Adrian Ulges Csaba Domokos Cercel Constantin Team:

Transcript of DOOR ENTRY SYSTEM Alina Dinca László Papp Adrian Ulges Csaba Domokos Cercel Constantin Team:

DOOR ENTRY SYSTEM

Alina Dinca

László Papp

Adrian Ulges

Csaba Domokos

Cercel Constantin

Team:

What is Project 9 about?

Name: Door entry system – feature analysis of a face using point separation

Input: images of several faces Operation: Identify key points (eyes, end of nose, mouth).

Measure distances and angles between these (for different orientations). Feed the results into a statistical analysis routine. Identify for unknown image most likely match.

Coding: C++, Matlab Remarks: difficulty quite hard

What we have Data base with grayscale pictures in the .pgm format

What we want to achieve

Locate the key points

Make a classification

algorithm for .pgm reader extract 64/64 keypoint cut-outs make an average (pattern) for each group

of cut-outs

Step 1 Locating key points

( )

transform the patterns .pgms with Fast Fourier Transformation transform the input image with Fast Fourier Transformation convolute the input image with each pattern to find the maximum transform them back from the Fourier space

Idea1. Using FFT => didn’t work!

FFT

FFT

*Response image

Inverse FFT

The formula for it is:

from {-1, 1}. If almost 1, then we have a match!! Get the maximum Slow algorithm (2½ minutes)

Idea 2. Similarity measure: correlation

maximumCorrelation image

-1

-1

1

1

2nd scaling1st scaling

1) Scaling the input and the average twice

2) Match in small image3) Find the match and scale back the match

4) Faster algorithm (6 seconds)

2nd scaling1st scaling

Idea 2. => Hierarchical Matching

--- A faster aproach ---

Input

Average

64/64

Evaluation:

10 pictures from the data base search eyes, noses, lips visual inspection Results

eye - 80% nose - 80% lip - 20%

Side knowledge about

keypoints?

use 20 key points from Data Base feature vectors: normalized coordinates (form a neuronal network)

use the nearest neighbour

Evaluation: - 1020 data records

- 510 training set

- 510 test set

- results: 98% recognition rate

Step 2 Make the classification

Acces denied Acces granted

New image

Training

-1-1

1

1

( )

THANK YOU…

… for your attention