Weighted Multi-scale Local Binary Pattern Histograms for Face Recognition

Transcription

1 Weighted ulti-scale Local Binary Pattern Histograms for Face Recognition Presenter: Olegs Nikisins Institute of Electronics and Computer Science, Riga, Latvia. Author: Olegs Nikisins Project: ultimodal biometric technology for safe and easy person authentication DP/...0/APIA/VIAA/ International Conference on Applied athematics and

2 Weighted ulti-scale Local Binary Pattern Histograms for Face Recognition 3 x 3 Neighborhood from input image LBP Local Binary Patterns Result of LBP transformation 03 International Conference on Applied athematics and

3 Weighted ulti-scale Local Binary Pattern Histograms for Face Recognition Calculate histogram for each region Stack histograms into a single feature vector 3 03 International Conference on Applied athematics and

4 Weighted ulti-scale Local Binary Pattern Histograms for Face Recognition Database Input image Recognition stage: compare LBP histogram of the input image with histograms from the database nearest neighbor classifier (NNC) is used (Example: Euclidean distance) 4 03 International Conference on Applied athematics and

5 Recognition stage class classification problem (for example, face detection) x x x x x x Class x x x x x x x o x x x x x o x o o o o o o o x x x x o o x x o o o o o oo o o o o o o o o o o o o o o Class ANN, SV can be used Lot s of training data for each class Number of classes not high Face recognition task Class + x + x Class 4 Class 5 x oo x Class x + Class 6 x + Class 3 not enough data for ANN, SV (!) few training samples per class Number of classes is high(000 FERET) Nearest Neighbour Classifier is usually used 5 03 International Conference on Applied athematics and

6 Is Nearest Neighbor Classifier (NNC) the best solution? Why NNC is usually used? Problem: Having plenty of classes / persons Only a few training examples per class / person Possible improvement Artificial Neural Networks Support Vector achines... Weighted Nearest Neighbor Classifier (WNNC) Utilize statistical information from all classes 6 03 International Conference on Applied athematics and

7 Proposed in two levels: feature and block weighting: LBP image A Blocks B Presented approach amplifies the features/blocks, which are more relevant for the recognition by adjusting the weights. Weights are determined in the learning process International Conference on Applied athematics and

8 Only TWO training examples per class are needed ( & ) d Euclidean distance between weighted histograms 8 03 International Conference on Applied athematics and

9 Only two training examples per class are needed ( & ) d Euclidean distance between weighted histograms d d d 33 intra-class distance d 9 03 International Conference on Applied athematics and

10 Only two training examples per class are needed ( & ) d Euclidean distance between weighted histograms d d 3 d d d 3 d 33 d d inter-class distance selected randomly: mini-batch 0 03 International Conference on Applied athematics and

11 Only two training examples per class are needed ( & ) d Euclidean distance between weighted histograms d d 3 d d d 3 d 33 d d d /d 3 - d /d - d 33 /d 3 - d /d - selected randomly: mini-batch 03 International Conference on Applied athematics and

12 Only two training examples per class are needed ( & ) d Euclidean distance between weighted histograms d d 3 d d d 3 d 33 d d INIIZE J(weights) - cost function d /d 3 - d /d - d 33 /d 3 - d /d - s s s s + selected randomly: mini-batch S - sigmoid 03 International Conference on Applied athematics and

13 d Euclidean distance between weighted histograms d d 3 d d d 3 d d 3 d 3 d 33 d d 3 d Learning data selection: Selected randomly at each iteration Select smallest interclass distances for all persons after each N iterations 3 03 International Conference on Applied athematics and

14 Learning data selection: Selected randomly at each iteration Select smallest interclass distances for all persons after each N=0 iterations 4 03 International Conference on Applied athematics and

15 Each face is described with N=6384 LBT parameters. Lets simplify the task: Lets consider that each face (class) is described with parameters and we have only 3 persons (classes) in the database International Conference on Applied athematics and

16 Before optimization: Precision 00% each class is described with parameters After optimization: Precision 00% each class is described with parameter Data compression with same precision 6 03 International Conference on Applied athematics and

17 Each face is described with N=6384 LBT parameters. Lets simplify the task: Lets consider that each face (class) is described with parameters and we have 5 persons (classes) in the database International Conference on Applied athematics and

18 Before optimization: Precision 60% each class is described with parameters After optimization: Precision 00% each class is described with parameter Recognition accuracy is improved 8 03 International Conference on Applied athematics and

19 Results on FERET face database Input face Block weights High Low IT and Harvard University research: Title: 9 important results regarding face recognition by humans. One of the facts was: "of the different facial features, eyebrows are among the most important for recognition International Conference on Applied athematics and

20 Results on FERET face database Number of persons in the database is almost 000 frontal face images per person are available fa and fb sets Optimal learning data SLBP SLBP + feature weighting SLBP + block weighting 96,8 % 98. % 99. % 0 03 International Conference on Applied athematics and

21 Thank You! 03 International Conference on Applied athematics and