Facial Emotion Recognition FER

1
Facial Emotion Recognition(FER)

• Jaromir Horejsi

2
Goal
3
Images

• The Japanese Female Facial Expression (JAFFE)
  Database
• The database contains 213 images of 7 facial
  expressions (6 basic facial expressions 1
  neutral) posed by 10 Japanese female models

4
Images
5
Image preprocessing

• The most information about facial expression are
  located in the following region

6
Image preprocessing

• Original image in application window

7
Image preprocessing

• Rotated image, eyes are lined up

8
Image preprocessing

• Cropped image

9
Image preprocessing

• Image normalized to size 96x128

10
Image preprocessing

• Histogram equalization

11
Image preprocessing

• Histogram equalization

12
Dimensionality reduction

• After preprocessing, an image has 96x12812288
  pixels
• 12288 pixels is too much information for
  follow-up classification -gt we need to reduce
  this information (keep only the most significant
  information)

13
Principal Component Analysis (PCA)

• is a technique used to reduce multidimensional
  data sets to lower dimensions for analysis
• searches for directions in the data that have
  largest variance and subsequently projects the
  data onto it

14
Principal Component Analysis (PCA)

• Original data

15
Principal Component Analysis (PCA)

• After reduction of one dimension

16
Linear Discriminant Analysis(LDA)

• finds the linear combination of features which
  best separate two or more classes of object or
  event

17
PCALDA

• Combination of both previously mentioned
  techniques
• Should overcome both PCA and LDA

18
Testing

• 1) Subset of 4 emotions (124 images)
• Two randomly chosen pictures of each person and
  each emotion were used for training (10 persons x
  4 emotions x 2 images 80)
• The remaining pictures were used for testing (44
  images)

19
Testing

• 2) The whole database - 7 emotions (213 images)
• Two randomly chosen pictures of each person and
  each emotion were used for training (10 persons x
  7 emotions x 2 images 140)
• The remaining pictures were used for testing (73
  images)

20
Classification

• K-Nearest Neighbor Classifier

21
Results

• PCA performance, 4 emotions

22
Results

• LDA performance, 4 emotions

23
Results

• PCALDA performance, 4 emotions

24
Results

• PCA performance, 7 emotions

25
Results

• LDA performance, 7 emotions

26
Results

• PCALDA performance, 7 emotions

27
Conclusion

• 4 emotions - the best results
• PCA 88.64
• LDA 95.45
• PCALDA 93.18

28
Conclusion

• 7 emotions - the best results
• PCA 80.82
• LDA 86.30
• PCALDA 89.04