• No results found

4.2 Multispectral palmprint recognition based on local binary pattern histogram

4.2.3 Experiments and Results

4.2.3.1 Protocol I

The aim of this experiment was to evaluate the recognition performance of the proposed technique, with a standard evaluation used protocol similar to those in previous studies [67, 111, 113] is used. The PolyU [121] multispectral palmprint database was used. A total of 500 subjects of each of the blue, green, red and NIR spectra were used in the first protocol. In the training phase, the images obtained in the first session were selected while the remaining images were used in the testing phase (6 training samples and 6 testing samples).

Experiments were conducted to compare the proposed approach with a number of state- of-the-art, including the techniques described elsewhere [113, 111, 67] (see section 3.8.1) where NFS is the nearest feature space classifier which used the distance between the test sample and the space spanned by the training samples from a class. The discrete wavelet transform (DWT) method was designed to extract local features from palmprints. Jing at al. [67] introduced the two-phase test sample representation (TPTSR) method. In the first phase, TPTSR represents the test sample as a linear combination of all the training samples and determines the M nearest neighbours of the test sample. In the second phase, the test sample is linearly represented with new coefficients, weighting the M nearest neighbours in a linear combination to classify the test sample.

As shown in Table 4.1, the LBP-HF+Gabor approach yields the highest classification rate of 98.02% for the blue spectrum. Moreover, it can be seen that, in the case of the green spectrum, the LBP-HF+Gabor method achieves the best result, with a rate of 98.37%. Consequently, there is an improvement of 0.35% compared to TPTSR, a 2.0% improvement compared to NFS and an improvement of 4.87% compared to DWT. In the case of the red spectrum, the results clearly show that the LBP-HF+Gabor approach achieves the best result, with a performance accuracy of 98.74%, representing a gain of 0.16-3.54% in relation to the other reported methods. With regard to the NIR spectrum, the LBP-HF+Gabor approach also obtains the best result, with a performance accuracy of 98.67%, which is also higher than those of the existing TPTSR, DWT and NFS methods. The ROC curve is drawn in Figure 4.1 to provide another biometric performance measure depicting the comparison between multispectral (blue, red, green and NIR) palmprint data sets.

Figure 4.1: ROC curves for the proposed method under the blue, green, red and NIR spectral using six images captured in the first session for the training set and six images captured in the second session for the testing set.

4.2.3.2 Protocol II

The proposed approach was implemented and tested on a similar palmprint database, com- paring the proposed LBP-HF+Gabor method to two existing methods, namely NFS [113] and RBF [132]. The proposed approach was evaluated using the multi-spectral palmprint database and followed the same evaluation given in previous studies [113, 132]. The training set consisted of the first three images of each of the red, green, blue and NIR spectra from the first session and the testing dataset consisted of the 6 palmprints of each of the blue, green, red and NIR spectra obtained from the second session.

Table 4.2: Comparison of recognition rates of the proposed method with the state-of-the- art methods for different spectra bands: blue, green, red and NIR. The computation of the recognition rate is obtained for a palmprint with 3 samples for training and 6 samples for testing in Protocol II.

Methods Recognition Rate (%)

Blue Green Red NIR

NFS [113] 95.10 92.87 95.40 95.63

RBF [132] 96.70 96.50 98.20 98.40

Proposed LBP-HF+Gabor 97.70 97.44 98.24 98.57

A comparative analysis of the proposed method was performed against two state-of-the- art methods, including the method proposed by Yong et al. [113], and on the radial basis

function (RBF) [132]. The former method has been introduced in protocol I. As previously mentioned in subsection 3.8.2. the RBF kernel function is used to optimize the feature space such that the samples from the same class are well clustered while the samples from different classes are pushed far away.

Figure 4.2: ROC curves for the proposed method under the blue, green, red and NIR spectral using three images captured in the first session for the training set and six images captured in the second session for the testing set.

The results shown in Table 4.2 clearly demonstrate the superiority the LBP-HF+Gabor approach in terms of robustness, in addition to its greater effectiveness compared to the other reported methods. The LBP-HF+Gabor approach offers attractive recognition per- formance rates of 97.44% to 98.57%, which are highlighted in bold in Table 4.2. It should be noted that the recognition rate using the LBP-HF+Gabor approach in relation to the blue spectrum yields an improvement of 1-2.6% compared to the RBF and NFS methods. In the case of the green spectrum, the LBP-HF+Gabor method achieves an accuracy of 97.44%, which is also higher than all other reported methods. Moreover, the LBP-HF+Gabor method also achieves better results of 98.24% and 98.57% respectively for the red and NIR spectra than the RBF and NFS methods. The ROC shown in Figure 4.2 plots the false acceptance rate versus the genuine acceptance rate for the multispectral palmprint data sets, represented by the red, green, blue and NIR spectra.

4.2.4 Discussion

The proposed method successfully captures discriminating information to provide better recognition performance using the KNN classifier. By looking at the results, it can be seen that the proposed methodology achieves outstanding recognition rates compared to other

existing approaches. The experiments conducted were able to achieve recognition rates of 97.44%-98.74% for the blue, green, red and NIR spectra as shown in tables 4.1 and 4.2.

Figure 4.3: Comparison of recognition rates of the proposed method with Protocols I and II for different spectra (blue, green, red and NIR)

It can also be observed from the tables that the recognition rate for the red illumination provides the highest performance compared to the NIR, green and blue spectra for the six training and six test images in Protocol I. The results obtained from the NIR spectrum have shown an improvement compared to the spectra with three training and six testing images in Protocol II. From the experimental results, there seems to be an improvement in the Protocol with six training images and six test images as shown in Figure 4.3. The red and NIR spectral are able to achieve better performance levels than blue and green in both protocols. As mentioned previously in section 3.7, the NIR and red spectral capture the palm lines and vein structures, as shown in Figure 2.11 in chapter 2; this helps in the comparison and classification of palms with similar palm lines [121].

4.3

Multi-feature analysis based on the shift binary pattern