On co-training online biometric classifiers
Summary (2 min read)
1. INTRODUCTION
- A biometric verification system typically uses a classifier to determine if the unlabeled probe data matches with the labeled gallery data.
- New information that can affect the biometric data distribution (e.g. match scores) is available from two fronts: (1) new subjects enrolling into the biometric system (labeled data) and (2) previously enrolled subjects interacting with the system and providing new probes (unlabeled data).
- Intuitively, ∙ labeled information from newly enrolled individuals can be used to update the classifier in incrementaldecremental learning mode, also known as online learning.
- The paper presents a framework for co-training biometric classifiers in an online manner.
2. Proposed Co-training Online Framework
- Mathematically, for a two classifier biometric verification system, the process is as follows.
- Every instance u𝑖 or u′𝑖 has two views, u𝑖 = {𝑥𝑖,1, 𝑥𝑖,2}; here 𝑥𝑖,1 and 𝑥𝑖,2 represent the match scores obtained from the two classifiers and the label 𝑧𝑖 ∈ {+1,−1} represents the genuine or impostor class.
- In online learning, classifiers 𝑐1 and 𝑐2 are updated for every incorrect prediction (i.e., when 𝑦𝑖 ∕= 𝑧𝑖) while no action is taken when the instances are correctly classified.
- Classifiers are co-trained for a given instance if one classifier confidently predicts the label of the instance while the other classifier is unsure of its prediction.
2.1. Online SVM Classifiers
- Since u𝑖 represents the two individual views , SVMs are trained individually for both the views using 𝑥𝑖,𝑗 where 𝑗 = 1, 2. is the mapping function used to map the data space to the feature space, and 𝐶 is the tradeoff parameter between the permissible error in the samples and the margin.
- SVM classifiers for each view/score are first trained on the initial enrolment training data 𝐷𝐿.
- SVM classifiers are then used to classify each of these match scores as genuine or impostor.
- The online learning algorithm to update the classifiers is described in Algorithm 1.
2.2. Co-training SVM Classifiers
- In biometrics, obtaining a large number of labeled examples is a difficult and expensive task.
- The two classifiers are first trained on an initial small labeled data set.
- Similarly, for an instance to be confident enough to lie in the impostor class, its distance from the decision hyperplane should be greater than the impostor threshold.
- Set of labeled training data 𝐷𝐿, set of unlabeled instances 𝐷𝑈 , where each instance u′ = (𝑥𝑖,1, 𝑥𝑖,2) represents two view/scores.
- The proposed co-training framework is illustrated in Figure 3 and described in Algorithm 2.
2.3. Co-training Online SVM Classifiers
- The online learning and co-training approaches are extended to propose a framework that simultaneously uses online learning and co-training to update the classifier using labeled and unlabeled data as and when they arrive.
- The classifiers are initially trained on a small labeled training data set.
- For every new user being enrolled in the system, online learning is used to update the classifiers using the labeled data generated during enrolment.
- During probe verification, whenever a user queries the system, co-training is used to update the classifiers using the unlabeled data.
3. Case Study: Multi-classifier Face Verification
- To evaluate the effectiveness of the proposed co-training framework, experiments are performed using a multiclassifier face verification application.
- Pointbased Speeded Up Robust Features (SURF) [6] and texturebased Uniform Circular Local Binary Pattern [3] are used as facial feature extractors along with 𝜒2 distance for matching.
- The final classification is obtained by combining the responses from the two updated classifiers using SVM fusion [11].
- To analyze the performance on a large database, images from multiple face databases are combined to create a heterogeneous face database of 1833 subjects.
- Though each constituent database has large number of images per subject, images exhibiting large pose (> 30 degree), extreme illumination conditions, and occlusion are ignored.
3.1. Experimental Protocol
- The experimental protocol is designed such that the classifiers are first trained on labeled training data and then variations due to new enrolments and probes are simultaneously learned using online learning and co-training.
- To evaluate the effectiveness of co-training, two experiments are performed.
- The cotraining is performed using the probes of all 1833 subjects and this experiment is termed as cotraining-1. –.
- In the second experiment, the classifiers are trained using all 1833 subjects in batch mode and co-training is performed using the probe images.
- The results are reported based on five-fold nonoverlapping random cross validation and verification accuracies are computed at 0.01% false accept rate (FAR).
3.2. Results and Analysis
- Figure 4 shows the Receiver Operating Characteristic (ROC) curves for the multi-classifier face verification system.
- The framework improves the performance by at least 0.54% compared to batch learning, online learning, and co-training.
- Co-training provides an improvement in verification accuracy over both batch learning and online learning because the classifiers trained on different scores update each other by providing pseudo labels for the instances where the other classifier makes an error.
- If the correlation between individual classifiers is high, the improvement due to co-training may be limited. ∙.
- For the proposed framework, classifier1 was updated on 34, 086 instances and classifier2 was updated on 42, 102 instances using co-training during probe verification.
Did you find this useful? Give us your feedback
Citations
119 citations
Cites background from "On co-training online biometric cla..."
...• Decision update Researchers are exploring the use of online or incremental learning approaches to improve the decision boundary of the classifiers even in deployment phase [32,33]....
[...]
111 citations
85 citations
68 citations
Cites methods from "On co-training online biometric cla..."
...Classifier update using co-training is explored by Bhatt et al. [38] where the biometric classifiers are updated using labeled as well as unlabeled instances....
[...]
51 citations
Cites methods from "On co-training online biometric cla..."
...Besides subject-specific incremental improvements, new quality-controlled data can also be employed to improve biometric models via online learning [80][68]....
[...]
References
13,011 citations
"On co-training online biometric cla..." refers methods in this paper
...2In our experiments, SURF and UCLBP had genuine Pearson’s correlation of 0.58 and impostor Pearson’s correlation of 0.46....
[...]
...UCLBP and SURF are used for facial feature extraction because they are fast, discriminating, rotation invariant, and robust to changes in gray level intensities due to illumination variations....
[...]
...Pointbased Speeded Up Robust Features (SURF) [6] and texturebased Uniform Circular Local Binary Pattern (UCLBP) [3] are used as facial feature extractors along with 𝜒2 distance for matching....
[...]
...Two SVM classifiers, one for SURF (classifier1) and another for UCLBP (classifier2), are trained to classify the scores as 𝑔𝑒𝑛𝑢𝑖𝑛𝑒 or 𝑖𝑚𝑝𝑜𝑠𝑡𝑜𝑟....
[...]
...Pointbased Speeded Up Robust Features (SURF) [6] and texturebased Uniform Circular Local Binary Pattern (UCLBP) [3] are used as facial feature extractors along with χ(2) distance for matching....
[...]
5,840 citations
"On co-training online biometric cla..." refers background in this paper
...Online learning [18] and co-training [7] are used to update the classifiers in real time and make them scalable....
[...]
...Blum and Mitchell [7] showed that two classifiers should have sufficient individual accuracy and should be conditionally independent of each other....
[...]
...In co-training, as proposed by Blum and Mitchell [7], two classifiers that are trained on separate views (features), co-train each other based on their confidence in predicting the labels....
[...]
5,563 citations
"On co-training online biometric cla..." refers methods in this paper
...2In our experiments, SURF and UCLBP had genuine Pearson’s correlation of 0.58 and impostor Pearson’s correlation of 0.46....
[...]
...UCLBP and SURF are used for facial feature extraction because they are fast, discriminating, rotation invariant, and robust to changes in gray level intensities due to illumination variations....
[...]
...Pointbased Speeded Up Robust Features (SURF) [6] and texturebased Uniform Circular Local Binary Pattern (UCLBP) [3] are used as facial feature extractors along with 𝜒2 distance for matching....
[...]
...Two SVM classifiers, one for SURF (classifier1) and another for UCLBP (classifier2), are trained to classify the scores as 𝑔𝑒𝑛𝑢𝑖𝑛𝑒 or 𝑖𝑚𝑝𝑜𝑠𝑡𝑜𝑟....
[...]
...Pointbased Speeded Up Robust Features (SURF) [6] and texturebased Uniform Circular Local Binary Pattern (UCLBP) [3] are used as facial feature extractors along with χ(2) distance for matching....
[...]
3,773 citations
Related Papers (5)
Frequently Asked Questions (15)
Q2. What future works have the authors mentioned in the paper "On co-training online biometric classifiers" ?
As future work, the proposed framework can be extended to different stages of a biometric system that require regular updates. The authors also plan to incorporate the quality of the given gallery-probe pair in computing the confidence of prediction rather than making a decision based only on the distance from the hyperplane.
Q3. What is the purpose of the proposed framework?
In the proposed framework, co-training is used to leverage the availability of multiple classifiers and unlabeled instances to update the decision boundaries of both the classifiers and account for the wide intra-class variations introduced by the probe set.
Q4. What is the purpose of the proposed co-training framework?
During probe verification, whenever a user queries the system, co-training is used to update the classifiers using the unlabeled data.
Q5. How many instances were used for classifier update?
For online learning, during enrolment, classifier1 was updated using 22,145instances and classifier2 was updated using 31,846 instances.
Q6. What is the main reason for the change in the score distribution of a biometric system?
New enrolments can lead to variations in genuine and impostor score distributions while probe images may introduce wide intra-class variations (due to temporal changes).
Q7. What is the purpose of retraining a biometric system?
To maintain the performance and to accommodate the variations caused due to new enrolments and probes, biometric systems generally require re-training.
Q8. What is the advantage of online SVM classifiers?
online SVM classifiers have a significant advantage of reduced re-training time using only the new sample points to update the decision boundary.
Q9. What is the way to update a classifier?
Since corresponding labels (“genuine” or “impostor”) are available during enrolment, classifier update using online learning can be viewed as a supervised learning approach.∙ unlabeled information obtained at probe level can be used to update the classifier using co-training.
Q10. What is the mapping function used to map the data space to the feature space?
is the mapping function used to map the data space to the feature space, and 𝐶 is the tradeoff parameter between the permissible error in the samples and the margin.
Q11. How many users can use online learning algorithms?
Kim et al. [14] have shown that online learning algorithms can be used for biometric score fusion in order to resolve the computational problems with increasing number of users.
Q12. What is the process to update a classifier?
Iterate: 𝑗= 1 to number of views (number of classifiers) Process: Train classifier 𝑐𝑗 on 𝑗𝑡ℎ views of 𝐷𝐿 for 𝑘 = 1 to 𝑁 doPredict labels: 𝑐𝑗(𝑥𝑖,𝑗) → 𝑦𝑖 if 𝑦𝑖 ∕= 𝑧𝑖 thenUpdate 𝑐𝑗 with labeled instance {𝑥𝑖,𝑗 ,𝑧𝑖} end ifend for End iterate Output: Updated classifier 𝑐1 and 𝑐2.
Q13. How many instances were updated using co-training?
For the proposed framework, classifier1 was updated on 34, 086 instances and classifier2 was updated on 42, 102 instances using co-training during probe verification.
Q14. What is the way to train a classifier?
In co-training, as proposed by Blum and Mitchell [7], two classifiers that are trained on separate views (features), co-train each other based on their confidence in predicting the labels.
Q15. How many sample points can be controlled by co-training?
By varying the confidence threshold for a classifier, the number of sample points on which co-training is performed can be controlled.