DATABASE

We first summarise the multi-spectral ocular database employed for gender classification based on the ocular region in this section (Figure 8.1). The database for this experimental evaluation is collected using our custom-built multi-spectral imaging sensor (Vetrekar et al., 2016) to acquire the images in eight narrow spectrum bands corresponding to 530nm, 590nm, 650/ш, 710/ш, 110 nm, 890nm, 950nm and 1000 nm spanning from VIS to NIR wavelength range. The acquired ocular database for gender classification consists of 104 unique ocular images from 52 subjects with a distribution of 32 male and 20 female participants.

As the goal of this study is to perform the gender classification when the ocular region is occluded with eyeglasses, we have collected multi-spectral images for the ocular region with and without the presence of glasses over two sessions. The first set of data collection comprises multi-spectral ocular instances when subjects are not wearing eyeglasses and in a simplified manner, we call this set of images by an acronym “Without-Glass” data. On similar lines, the second set of data collection comprises multi-spectral ocular instances when subjects are asked to wear eyeglasses and in a simplified manner, we call this set of images by an acronym “With-Glass” data. Table 8.2 presents the description of acronyms used for two different categories of data collection.

FIGURE 8.1 Multi-spectral ocular instances collected using eight narrow bands across visible and near-infra-red spectrum. Figure illustrates the sample ocular images for male and female class for two classes of “Without-Glass” and “With-Glass”.

TABLE 8.2

Description of Acronyms Used for Two Different Categories of Multi-Spectral Images Collected for Ocular Region

Further, each category of ocular image is captured in five instances in two different sessions, and the time difference between each of these sessions was 3-4 weeks. The sample ocular images were collected at a stand-off distance of one meter between the multi-spectral imaging sensor and subject in a controlled indoor illumination conditions. Proper care has been exercised while selecting the eyeglasses for this experimental data collection, in order to avoid any kind of reflections due to the illuminations from the indoor lighting. Under each category of ocular instances for male and female, a total of 8320 ocular instances [corresponding to (64 Male Ocular Instances x 2 Sessions x 5 Samples x 8 Bands) + (40 Female Ocular Instances x 2 Sessions x 5 Samples x 8 Bands) = 8320 total ocular instances] are obtained. The detailed summary representing the total number of ocular instances corresponding to two different categories across male and female class is illustrated in Table 8.3.

Further, the database employed in this work has following unique points compared to existing database employed in the ocular biometric for gender classification:

• • The database available in the public domain are mainly in VIS spectrum or NIR spectrum, while these databases, especially VIS spectrum, do not contain the distinct information due to the integration process that exist in the formation of image over broad spectrum [except our recent work (Raja et al., 2020)]. However, the database presented in this work addresses these limitations by capturing the ocular images in narrow spectrum bands in discrete and disjoint manner to obtain intrinsic features.
• • Further, the use of narrow spectrum band in discrete and disjoint manner allows to extend the inherent characteristic details in the spatial spectral domain without any redundancy in the data.
• • Majority of previous works have used the facial images to crop the ocular region for gender classification, while we have build our database specially for ocular biometric only.
• • Also most of the work uses the eyebrow information in their database, however this work eliminates the effect of eyebrow in the database to avoid any kind of biasing in the proposed work

TABLE 8.3

Summary Illustrating the Total Number of Ocular Instances Collected for Male and Female Using Multi-Spectral Imaging

8.3.1 Data Preprocessing

The sample ocular instances collected for “With-Glass” category of data collection consists of eyeglass. To avoid any bias in just classifying the frames alone, we process the images to remove frames from the ocular images by cropping the ocular region, as shown in Figure 8.2. As the original multi-spectral images collected for the ocular region is of spatial resolution of 1280 x 1024 pixels, the images result in 270 x 160 pixels after cropping. On similar lines, the ocular sample images corresponding to “Without-Glass” category of database is also preprocessed for the analysis. Figure 8.2 illustrates the cropping of male and female ocular regions corresponding to the left and right eye region. Further, these sample multi-spectral images collected in eight different wavelengths, hence to maintain the similar illumination across all the eight spectrum bands, we have employed histogram enhancement technique to enhance the contrast uniformly across individual spectral bands.