A hybrid segmentation scheme for improved dorsal hand vein recognition
The dorsal hand vein (DHV) pattern is a highly secured biometric system that is significantly used in many applications due to its uniqueness. Although it is a safe and secure means for biometric identification, accurate recognition of vein patterns for this application remains challenging. To solve...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English English English |
| Published: |
2023
|
| Subjects: | |
| Online Access: | http://eprints.uthm.edu.my/10996/1/24p%20WAHEED%20ALI%20LAGHARI.pdf http://eprints.uthm.edu.my/10996/2/WAHEED%20ALI%20LAGHARI%20COPYRIGHT%20DECLARATION.pdf http://eprints.uthm.edu.my/10996/3/WAHEED%20ALI%20LAGHARI%20WATERMARK.pdf |
| Summary: | The dorsal hand vein (DHV) pattern is a highly secured biometric system that is significantly used in many applications due to its uniqueness. Although it is a safe and secure means for biometric identification, accurate recognition of vein patterns for this application remains challenging. To solve the issue, various machine learning (ML) and deep learning (DL) techniques were employed in the past to identify DHV correctly. A hybrid ML and DL strategy are adopted in this study. An automatic segmentation technique designed based on the histogram, thresholding and morphological operations is proposed to overcome the shortcomings of manual segmentation. The Bosphorus database is used for demonstration. While the first set of the experiment used the original segmented dataset, the second combines the original dataset with the augmented images generated using the combinations of rotation transformations (i.e., [30˚ -30˚] and [50˚ -50˚]) and flipping. The results comparing the performance of AlexNet, which is used as the baseline, revealed a considerable difference between the outputs trained using manual and automatically segmented datasets with a classification accuracy of 87.5% and 76.5 %. This difference in accuracy is significantly reduced to 4 % with the augmentation methods i.e., 91.5 % and 88 %. Interestingly, the inclusion of augmentation does not increase the performance in the manual likely because the existing data is sufficient for the model to learn all core features. The proposed segmented set with augmentation is further supported by the good classification performance of GoogleNet and ResNet-18. The mean and standard deviation of AlexNet, GoogleNet and ResNet-18 in their classification accuracy, sensitivity, and specificity are given by 99.79±0.098 %, 89.5±4.92 %, and 99.89±0.05 %. The ResNet-18 achieved superior performance with less training time than GoogleNet on the DHV dataset, which can be attributed to its capacity to address the network degradation issue. This work recommends the proposed framework and a deep model with skip connections, such as ResNet-18 for use in recognizing DHV patterns for future authentication research and system development |
|---|