F. Simulations
In order to illustrate the capabilities of linear correlation filters for facial recognition, the captured datasets were used to develop a classifier based on facial features using MACE and HBCOM filters. By ROC analysis, the optimal threshold values for the classifiers were computed and the corresponding performance indicators were calculated. The results are presented bellow.
The simulations for this filter were carried out using the corresponding protocol described in the Readme. The following parameters were used when building the filters:
| Subject | Ref. Imag. | Num. Images |
|---|---|---|
| A | 2 | 2 |
| D | 12 | 2 |
| G | 102 | 2 |
| N | 39 | 2 |
The following ROC curves were obtained carrying out the process described in the methodology. The figures correspond to A (upper left), D (upper right), G (bottom left) and N (bottom right).
From the plots above, the optimal threshold metric was computed, and also the performance indicators associated to the classifier algorithm.
| Subject | Threshold | TPR | FPR | FNR | TNR |
|---|---|---|---|---|---|
| A | 11.8633 | 0.8844 | 0.0854 | 0.1156 | 0.9146 |
| D | 28.4959 | 0.9548 | 0.0241 | 0.0452 | 0.9759 |
| N | 9.7330 | 0.6432 | 0.3849 | 0.3568 | 0.6151 |
| G | 16.4256 | 0.9950 | 0.0050 | 0.0050 | 0.9950 |
The simulations for this filter were carried out using the corresponding protocol described in the Readme. The following parameters were used when building the filters:
| Subject | Ref. Imag. | Num. Images |
|---|---|---|
| A | 46 | 4 |
| D | 1 | 4 |
| N | 165 | 4 |
| G | 40 | 4 |
The following ROC curves were obtained carrying out the process described in the methodology. The figures correspond to A (upper left), D (upper right), G (bottom left) and N (bottom right).
From the plots above, the optimal threshold metric was computed, and also the performance indicators associated to the classifier algorithm.
| Subject | Threshold | TPR | FPR | FNR | TNR |
|---|---|---|---|---|---|
| A | 8.1815 | 1.000 | 0.000 | 0.0000 | 1.0000 |
| D | 8.5902 | 1.000 | 0.000 | 0.000 | 1.0000 |
| N | 6.1542 | 0.7538 | 0.2774 | 0.2462 | 0.7226 |
| G | 7.7166 | 0.9648 | 0.0191 | 0.0352 | 0.9809 |
Based on these results, is clear that both filters show an acceptable discrimination most of the times since the false positive rate is significantly small in most cases and correspondingly, true positive rate values are all above 80%, which is the ideal range of accuracy set out as the ultimate goal. Furthermore, the HBCOM filter designed for this work exhibits a remarkable performance, as can be seen by the indicators. Consequently, the HBCOM filter is presented as the utmost result of the project in terms of reliability and efficiency. The results obtained for subject N are not consistent with the rest of the subjects. This might be due to the fact that the images on the corresponding database included background objects since subject's face has very different dimensions when compared to other subjects. In the data acquisition algorithm, this is a failure but compromise between several subject database acquisition and extraction of the region of interest has to be made. However, other 2 data sets were tested using the project protocol, and results similar to subjects A, D and G were obtained, thus showing the high influence that a proper database has in non-segmentation biometric recognition. Those results are not included for publish permit from the subjects was not given.
The overall performance of the correlation algorithms is impressive, and the results show that for variations in facial expression, the technique produces EERs less than 10% for most datasets. The preprocessing techniques included proved to be valuable for increasing subject discrimination, specially for MACE filters. The project shows that proper database acquisition is fundamental. Further work might include variations of the facial plane in the datasets.