*4.2. Tests of the Experimental System Speed*

Time taken to evaluate a tested image is between 0.6 and 0.9 seconds. In terms of calculations, it was somewhat difficult to derive the algorithm for the extraction of biometric characteristics. Table 4 summarizes the processing speed of the individual program steps and the corresponding image evaluation.

#### *4.3. Results*

FMR (False Match Rate), FNMR (False Non-Match Rate), and EER functions were utilized in evaluating the biometric system. These functions specify the faults frequency of the system. First, the comparison of the individual metrics for the data gathered from participating persons were compared for single-biometric and multi-biometric systems. The threshold value was chosen to be 0.5 [-] and the boundary of the minimum interval varied from 0 to 0.5 [-] (Figure 13). For the metrics normalization, the previously min-max method was applied. The graphs presented in Figures 14 and 15 capture the test results of the two different metrics for hand geometry. Graphs in Figures 16 and 17 capture the test results of the two different metrics for the bloodstream.

In the next part of the testing, the best metric for hand geometry (Hamming metric) as well as the best metric for the bloodstream (MHV) are selected. With these metrics, the multi-biometric system was created and the performance was tested against the data described above (Table 5). The results are shown on the graph in Figure 13. As expected, results showed that the multi-biometric system performed better in comparison to a single biometric system. The EER value for the multi-biometric system was found to be half of what is obtained from a single bloodstream biometry. Moreover, the progress of the FNMR fault was smoothened and partially reduced.

**Figure 13.** EER for united metrics of Hamming and MHV.



**Figure 14.** EER for Euclidean distance.

**Figure 15.** EER for Hamming distance.

**Figure 16.** EER for HV.

**Figure 17.** EER for MHV.

#### **5. Discussion**

The database developed for this study was built using 280 images (finger view and palm view, vascular images and back of the hand) from 40 individuals, i.e., seven images per subject. Two of the images generated per person (Total of 80 images) was used to train available researchers, so that they get familiar with the systems. Nevertheless, these images were carefully derived so that they were also useful for analysis. In testing our verification technique, each of the generated biometric data is technically subdivided (partitioned) into 2 × 50, 1 × 80, and 1 × 100 samples. Individual scores (from each uni-modal) were then summed up to arrive at the overall score. The equal error rate of the blood stream (11%) and hand geometry consolidated with the blood stream (5%). This approach is dependent on score-level fusion with the individual. Score results were normalized to achieve the stated percentages.

In addition to the samples used for training, extracted characteristics per scanned hand is related to how the proposed multimodal system works. The more the number of scanned hands, the more the extracted features and the better the behavior of the proposed multi-modal systems in terms of its performance. The implication of this is that the system witnesses some form of increased computational cost, which is similar to the finding of Reference [68]. As such, the number of extracted characteristics per scanned hand is checked against the EER. As noted by Reference [68], there is a possibility for an increment in the cost of computation in terms of the overall number of extracted characteristics. For example, this cost covers the time needed to input the first set of features into the database, which, thereby, trains the system, in addition to taking it for verification. Verification time is dependent on the speed of the computer (specification) used, as well as the algorithm for the experiment. There are a number of systems with rapid verification, segmentation, and feature extraction times. Some of these features utilize multi-dimensional vectors with more than two biometric approaches. In general, a high number of extracted characteristics imply longer processing times.

A number of challenges were noticed in the course of the experimentation with the proposed multi-biometric device. However, many of these challenges can be resolved at the point of use except for the issue posed by the extraction algorithm, which causes a lengthy time of verification. Algorithms of the individual metrics (hand geometry and blood stream analysis) demonstrates specific speed and precision issues that were noticed during experimentation, and may have significant effects on results. For instance, the low contrast produced by the bloodstream image during the investigation proved to be a limiting factor. There are also problems of discontinuation of veins and shadow produced by the human skin. This shadow image is confused to be another vein. Such small contrasts are most likely caused by a large distance between the lighting of the scene and the scanned hand. Camera

resolution of 640 × 480 pixels is another factor that seems to affect results negatively. This produced very small differences in measured widths of the fingers among persons. This problem can, however, be resolved by using cameras of higher resolution to produce more precise evaluation and differentiation results. However, this was not verified in this study due to time constraint. Differentiation and evaluation of individual metrics is done via frequency comparison of the double type faults (false identification) with the EER values. The lower the EER value, the higher the level of precision of the given metrics. The least obtained value was measured for modified Hausdorff distance metric (MHV), which yielded an EER value that was 11%. Furthermore, the best metrics for the given biometric characteristics (Hamming distance and MHV) were chosen and consolidated into the final evaluation, which results in a multi-biometric model with an EER value of 5%. This implies that, as expected, metrics consolidation meant improved system precision.

The biggest problem of the proposed multi-biometric system could be the unwanted influences from the surrounding. For instance, direct sunlight would influence the intensity of the scanned image. A change in features of the platform on which the scanned hand is laid can also influence the results. Lastly, the user-friendliness of the hardware for hand scanning would need to be improved in order to make the usage of the system faster and more comfortable.
