EXPERIMENTS AND RESULTS

This section presents in detail the evaluation protocol employed for this study and related results on gender classification based on using ocular instances. The classification accuracy is computed on the multi-spectral ocular images collected in eight narrow spectrum bands across the VIS and NIR wavelength range. The goal of this work is to examine the influence of wearing eyeglasses on the performance accuracy of ocular gender classification. To present the robustness in the classification accuracy, we present the results based on our proposed multi-spectral ocular gender classification approach. Specifically, we select four discriminative ocular band images corresponding to the highest entropy value and process independently using the bank of Gabor filters to extract local and global features. Finally, the Gabor feature vectors corresponding to selected spectral band images are concatenated to learn the classifier model using ProCRC for gender classification problem. We present the extensive set of classification results in the form of average classification accuracy on the larger dataset of 16640 ocular images collected using multi-spectral imaging sensor. To present the average classification accuracy, we have performed 10-fold cross-validation experiment to randomly select the ocular sample images for training and testing set in a disjoint manner without overlap. The results are presented using our proposed method, and comparison is provided against five different state-of-the-art methods to present its significance.

Experimental Evaluation Protocol

In this section of the chapter, we present the experimental evaluation protocol employed in this work for ocular gender classification. To present the gender classification using multi-spectral images collected for ocular instances, we present our experimental evaluation protocol where images are partitioned into training and testing set. The training set comprises an equal number of 20 male and 20 female ocular instances including their samples from “Without-Glass” data (Table 8.4). The total number of sample images in the training set consists of 3200 ocular images [corresponding to (20 Male ocular instances x 2 Sessions x 5 Samples x 8 Bands) + (20 Female ocular instances x 2 Sessions x 5 Samples x 8 Bands) = 3200 images]. The testing set comprises 44 male and 20 female ocular instances including their samples from “With-Glass” data. The total number of sample images in the testing set consists of 5,120 ocular images [corresponding to (44 Male ocular instances 2 Sessions 5 Samples 8 Bands) + (20 Female ocular instances x 2 Sessions x 5 Samples x 8 Bands) = 5120 images].

To present the significance of this work, we have employed two sets of experimental evaluation. Evaluation 1 corresponds to Without-Glass v/s Without-Glass, and Evaluation 2 corresponds to Without-Glass v/s With-Glass gender classification. The details of experimental results related to each of these evaluations are provided in the next sections. Further to present the fair comparison with our proposed approach, we have compared the classification accuracy with the performance of each individual spectral bands and across three different fusion methods. In the case of fusion, we have employed three different fusion methods corresponding to IMF (Li, Kang, Hu, & Yang, 2013), GFF (Li, Kang, & Hu, 2013), and 2-Discrete Wavelet Transform (DCT) (Amolins et al., 2007) to present our results. The gender classification results obtained across individual spectral bands, and the fusion of bands are performed independently using five different feature extraction algorithms such as LBP (Ojala et ah, 2002), LPQ (Ojansivu & Heikkil "a, 2008), HOG (Dalai & Triggs, 2005), GIST (Oliva & Torralba, 2001), and BSIF (Kannala & Rahtu, 2012). The performance evaluation results are obtained by processing independently these

TABLE 8.4

Experimental Evaluation Protocol Summarising the Total Number of Multi- spectral Ocular Instances Partitioned Under Training and Testing Set

Training Set

Database

Male

Female

Ocular

Instances

Sessions

Samples

Band

Ocular

Instances

Sessions

Samples

Band

Without-Glass

20

2

5

8

20

2

5

8

Total sample images

1.600

1.600

With-Glass

44

2

5

8

18

2

5

8

Total sample images

3.520

1.600

five different feature descriptor methods along with SVM classifier (Raghavendra et al., 2018; Vetrekar et al., 2017a, 2017b). Use of these feature extraction methods along with SVM classifier has recently been used in gender classification studies conducted on multi-spectral imaging data.

 
Source
< Prev   CONTENTS   Source   Next >