Building a robust facial recognition system based … Presentation... · BUILDING A ROBUST FACIAL...

Preview:

Citation preview

End of Year Presentation

Principle Investigator: David Pilkington

Supervisor: James Connan

BUILDING A ROBUST FACIAL RECOGNITION SYSTEM BASED ON GENERIC TOOLS

CONTENTS • Introduction

• Recap

• Changes to image database

• Testing Framework

• Benchmark System

• Design

• Testing Results

• Weaknesses and Proposed Solutions

• Developed System

• Design

• Implementation and Encountered Problems

• Experimental Results

• Concluding Thoughts and Possible Future Work

2

RECAP • Makes use of the EmguCV package (C# wrapper for OpenCV)

• Logitech 300 webcam (640 x 480)

• Benchmark System to establish baseline performance

• Identify weaknesses

• Propose and implement solutions

3

CHANGES TO THE IMAGE DATABASE • Comprised of 50 images

• 10 individuals with 5 images each

• Captured using the tools provided by the package

• Stored as .png files due to lossless compression (640 x 480)

• Varying backgrounds, lighting conditions, sexes, race

4

TESTING FRAMEWORK

5

• Consists of 5 tests:

Test Description 1 1 training image 2 2 training images 3 3 training images 4 4 training images 5 5 training images

BENCHMARK SYSTEM

6

DESIGN

Input Image EigenObjectRe

cogniser Identified

Face

7

training

TEST RESULTS

8

Test Recognition Rate 1 40% (s2, s3) 2 40% (s2, s3) 3 40% (s2, s3) 4 40% (s2, s3) 5 40% (s2, s3)

IDENTIFIED WEAKNESSES

• Lower recognition rate than we would like (40%)

• No ability to reject an image

• Highly afftected by spurious background noise

• Highly affected by varying lighting

9

PROPOSED SOLUTION

10

• Make use of face detection to crop image

• Make use of skin segmentation to further reduce background noise

• Use histogram equalisation to reduce light variability

• Introduce a threshold to the recongniser to allow the rejection of images

DEVELOPED SYSTEM

11

DESIGN

Input Image Face Detector Skin Segmentation

Eigen Recogniser

Identified Face

Image Rejected

12

training

FACE DETECTOR

13

SKIN SEGMENTATION: DESIGN

14

Input Image Nose Detector Colour Modelling Segmentation Segmented

Image

SKIN SEGMENTOR : NOSE IDENTIFICATION

15

SKIN SEGMENTOR: SKIN MODELING

16

SKIN SEGMENTOR: SEGMENTATION

17

EIGEN RECOGNISER

18

TRAINING

• Makes use of the eigenfaces technique which is base on PCA

• Receive training images and labels

• Create the PCA subspace

• Compute eigenvectors for training images

• Project the training images onto the PCA subspace to obtain eigenvalues for images

RECOGNITION

• Project test image onto the PCA subspace and obtain eigenvalues

• Find nearest neighbour by shortest Euclidean Distance

EXPERIMENTAL RESULTS

19

Test Recognition Rate 1 60% (s1, s2, s4) 2 60% (s1, s2, s4) 3 60%(s1, s2, s4) 4 80% (s1, s2, s3, s4) 5 80% (s1, s2, s3, s4)

EXPERIMENTAL RESULTS : COMPARISON

Euclidean Distance to Nearest Neighbour

20

EXPERIMENTAL RESULTS : COMPARISON

21

VS

Benchmark System Developed System

EXPERIMENTAL RESULTS : COMPARISON

22

Database Load Time

CONCLUDING THOUGHTS AND POSSIBLE FUTURE WORK

Conclusions • Background noise reduction has proved to be effective • Light sensitivity is still a problem • Thresholding was temperamental • Scalability problem due to image processes

Possible Future Extensions • Expand image data set • Comparison between recognition techniques • Effects of glasses, beards, etc. • Image verification

23

PROUDLY BROUGHT TO YOU BY …

24

Recommended