Evaluation 1: Without-Glass v/s Without-Glass

In this section of the chapter, we present in detail the experimental evaluation results, when the multi-spectral ocular instances corresponding to “Without-Glass” category of data are employed to learn the two class model during training, and comparison is performed against the ocular multi-spectral images corresponding to the same category of data, i.e., “Without-Glass” during testing, based on our proposed approach discussed in this chapter (refer Section 8.4) for gender classification. The purpose of this set of evaluation is to present the benchmark results when the same categories of data are learned and tested. The classification accuracy results obtained based on the proposed method is compared across the performance accuracy of individual bands and the fusion of bands. The performance of individual bands and fusion of bands is carried out for gender classification using five state-of-the-art methods such as LBP-SVM, LPQ-SVM. HOG-SVM. GIST-SVM, and BSIF-SVM. Based on the experimental evaluation protocol, we first present the results based on individual spectral bands (Section 8.5.2.1) and fusion of bands (Section 8.5.2.2) in the following sections.

Individual Band Comparison

We present in this section the performance analysis of individual spectral bands for ocular gender classification. Table 8.5 tabulates the average gender classification accuracy after 10-fold cross-validation, and Figure 8.4 illustrates the mean and variance plot describing the classification accuracy of individual spectral bands. The overall results have shown the reasonable average classification accuracy across the

TABLE 8.5

Average Gender Classification Accuracy (in %) across Individual Bands and Proposed Method, When Training Ocular Sample Images Belongs to Without-Glass and Testing Ocular Sample Images Belongs to Without-Glass Category of Data

Spectral Bands

Algorithm

530 nm

590 nm

650 nm

710nm

770 nm

890 nm

950nm

1,000nm

LBP-SVM

52.65

56.74

58.40

62.13

61.19

63.45

58.37

60.64

• LPQ-SVM

58.82

61.03

63.37

68.43

70.40

68.68

62.41

65.82

• HOG-SVM

53.27

54.80

70.57

69.07

66.53

67.26

60.23

62.21

•GIST-SVM

68.97

69.58

74.66

72.00

73.66

74.39

71.37

67.27

•BSIF-SVM

61.97

64.74

71.22

72.80

74.79

73.81

71.78

68.05

• Proposed approach

75.72

Average classification accuracy (%) illustrated in terms of mean and variance plot for Without-Glass v/s Without-Glass evaluation for gender prediction

FIGURE 8.4 Average classification accuracy (%) illustrated in terms of mean and variance plot for Without-Glass v/s Without-Glass evaluation for gender prediction.

individual spectral bands for gender prediction. Further, based on the evaluation results obtained for this set, we summarise our major observations as below:

  • • The highest average gender classification accuracy obtained across the individual band is 74.79% for 710 nm spectrum band using BSIF along with SVM classifier (BSIF-SVM). On the other hand, the lowest average gender classification accuracy obtained across the individual band is 52.65% for 530 nm spectrum band using LBP along with SVM classifier (LBP-SVM).
  • • Of the eight spectral bands employed in this work, bands such as 650 nm, 710nm, 770/ш, 890nm, and 950nm demonstrated consistently an higher classification accuracy for most of the algorithms used, while the bands such as 530nm, 590nm, and lOOOnm have indicated lower performance using state-of-the-art approaches. The better performance across 650 nm, 710wn, 770/m;, 890nm, and 950nm could be attributed to better Signal-To- Noise Ratio, as compared to the other spectral bands. However, their performance can be improved by using robust algorithms such as BSIF-SVM and GIST-SVM, as seen from the Table 8.5.
  • • Among the five feature descriptor algorithms used in this chapter, BSIF and GIST have indicated the highest performance accuracy compared to other algorithms such as LBP, LPQ, and HOG used in this work. The same can be very well observed from the mean and variance plot illustrated in Figure 8.4.

Fused Band Comparison

To provide the significance of individual band for classification accuracy, we also present the evaluation across three different fusion methods such as IMF, GFF, and 2-DWT for gender prediction. All the three methods were then analysed for the robustness by independently using five different feature extraction methods described above, followed by SVM classifier. Table 8.6 presents the average gender classification accuracy after 10-fold cross-validation. Figure 8.5 illustrates the mean and variance plot describing the classification accuracy of three different fusion methods using five different feature descriptor methods. A similar observation was

TABLE 8.6

Average Gender Classification Accuracy (in %) across Fusion of Bands and Proposed Method, When Training Ocular Samples Belongs to Without-Glass and Testing Samples Belongs to Without-Glass Category

Fusion Method

Algorithm

LBP-SVM

LPQ-SVM

HOG-SVM

GIST-SVM

BSIF-SVM

Proposed Approach

IMF

57.97

60.21

58.81

69.57

64.39

GFF

58

58.38

59.65

71.-53

64.96

75.72

2-DWT

59.78

68.36

73.57

72.65

72.88

Average classification accuracy

FIGURE 8.5 Average classification accuracy (%) illustrated in terms of mean and variance plot for three different fusion methods such as IMF. GFF, and 2-DWT. From the figure, the results corresponding to (a), (c), (e), represents the classification accuracy related to Without- Glass v/s Without-Glass evaluation and the results corresponding to (b), (d), and (f), represents the classification accuracy related to Without-Glass v/s With Glass evaluation.

made in terms of average classification across the fusion methods in comparison with the individual band performance, as illustrated in Table 8.5. Further, based on the results obtained, we summarise our specific observations as below:

  • • The highest average classification accuracy of 73.57% is obtained with the 2-DFT fusion method using the HOG-SVM algorithm, while the lowest average classification accuracy of 57.97% is obtained for the IMF fusion method using the LBP-SVM algorithm.
  • • Out of three different fusion approaches employed in this work, spectral band fusion based on 2-DFT demonstrates the better performance accuracy across all the five different state-of-the-art feature extraction methods. On the other hand, the fusion methods such as GFF, IMF have shown slightly poor classification accuracy compared to 2-DFT, but with the help of robust algorithms such as GIST, BSIF, their results are also comparable with 2-DFT.

Based on the benchmark results obtained in the above subsections for individual spectral bands and fusion of eight spectral bands, we can present the classification accuracy results based on our proposed approach. Tables 8.5 and 8.6 illustrate the average ocular gender classification accuracy, and Figures 8.4 and 8.5 illustrate the mean-variance plot describing the performance analysis of our proposed approach in comparison with individual spectral band and fusion of spectral bands performed using state-of-the-art gender classification techniques. The proposed approach has outperformed the state-of-the-art feature descriptor methods employed in this work for gender classification. Specifically, the new approach used for gender classification has obtained a maximum of 75.72% average classification accuracy compared to individual band and fusion of bands performance, as seen from Figures 8.4 and 8.5, respectively.

 
Source
< Prev   CONTENTS   Source   Next >