Hybrid ear recognition framework based on passive human identification
<p>Current identification of passive detection has been attention in the modern world due to the</p><p>system's robustness as an ear recognition framework based on a multiclassifier and attempt to</p><p>create user patterns v...
Saved in:
| Main Author: | |
|---|---|
| Format: | thesis |
| Language: | eng |
| Published: |
2022
|
| Subjects: | |
| Online Access: | https://ir.upsi.edu.my/detailsg.php?det=9579 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | <p>Current identification of passive detection has been attention in the modern world due to the</p><p>system's robustness as an ear recognition framework based on a multiclassifier and attempt to</p><p>create user patterns via extracted features from ear images, which have unique individual</p><p>identities. The collected features from the ear intersection points and the angles bounded</p><p>between curves using different descriptors and classifiers are considered unique information</p><p>used to generate unique features. The proposed framework commenced with the extraction of</p><p>eight sets of features (LBP, BSIF, LPQ, RILPQ, POEM, HOG, DSIFT, and Gabor) from 2D</p><p>ear images. Subsequently, ELM and SVM classifiers were trained on each set of features.</p><p>Seven combination rules (MR, AR, GWAR, ICWAR, Borda, DS, and AV (GWAR, Borda,</p><p>DS)) were utilized to acquire a total of 16 classifiers. Also, two optimization rules; genetic</p><p>algorithm and brute force were proposed for accuracy enhancement. The AWE and the USTB</p><p>datasets were utilized in the development, evaluation, and validation of an ear recognition</p><p>framework dataset. So, some vulnerabilities are observed in datasets and all challenges for ear</p><p>biometrics. The research findings showed that combining classifiers using different sets of</p><p>features yields better performance compared to using individual classifiers. However, using</p><p>one classifier or limited number is not enough to solve the problem of ear recognition with</p><p>different challenges such as Pose, Occlusion, Illumination, Blurry image, Rotation, Lighting,</p><p>Scale, and Translation. The validation of such a framework using the AWE dataset showed that</p><p>the SVM and ELM in combination with modern descriptors managed to enhance the</p><p>recognition. Rank-1 accuracy also reached 99% with Genetic Algorithm optimization, and 98%</p><p>with brute-force AR and brute-force GWAR. These results are compared to other results in the</p><p>literature and found to be superior. In conclusion, the main findings showed that the proposed</p><p>framework consisting of two classifiers SVM and ELM trained with selected features and the</p><p>combination rules managed to attain higher accuracy in-ear recognition compared with</p><p>previous studies. This ear recognition framework is a major step towards the recognition of</p><p>individuals from ears in real-world conditions. This study implies that the proposed ear</p><p>recognition framework based on ELM and SVM classifiers with combination and optimization</p><p>rules can be utilized to improve the effectiveness of passive human recognition where security</p><p>is of utmost importance.</p> |
|---|