The hyperspectral imaging acquisition system consists of two halogen lamps made by Osram, Inc., one charge coupled device (CCD) camera produced by Cooke, Inc., and one liquid crystal tunable filter manufactured by Meadowlark, Inc. The cost of the setup is approximately USD 6,500.00. The prototype of this acquisition system is illustrated in Figure 1.2. The CCD camera is placed in the middle with one halogen lamp on either side. The halogen lamps produce both visible light and NIR with spectra ranging from 520 to 1,040 nm. The light from the two halogen lamps irradiates on the palm or dorsal hand, and then reflects to the camera sensor for capturing images. A tunable filter is settled ahead of the camera lens and allows a single band to pass through its settings. To obtain stable spectral images, 10nm is set as the spectral distance in the tunable filter. Therefore, this hyperspectral

Schematic of our designed hyperspectral imaging device

FIGURE 1.2 Schematic of our designed hyperspectral imaging device.

Hyperspectral palm (the upper row) and dorsal hand (the lower row) samples

FIGURE 1.3 Hyperspectral palm (the upper row) and dorsal hand (the lower row) samples.

imagery acquisition system contains 53 bands in the range of 520-1,040nm with 10 nm intervals.

Each volunteer was asked to grasp a prop making a fist when capturing hisher dorsal hand images. Contrary to an open hand, a closed dorsal hand makes the vascular network more visible achieving discriminant feature exaction. For the palmprint, each individual placed his/her hand on a plate with pegs to somewhat fix their hand, while a cutout was made to expose the palm. Examples of hyperspectral palm and dorsal hand images captured using the designed apparatus are shown in Figure 1.3.


First, the ROI detection algorithms for hyperspectral palmprint images and dorsal hand vein images are introduced, respectively. Afterwards, several widely used patterns are presented for feature extraction. At last, a feature fusion strategy is proposed for multimodal recognition of hyperspectral palmprint and dorsal hand vein.

ROI Extraction

Hyperspectral Palmprint ROI Extraction

It is necessary to conduct ROI extraction from the palm image, due to the fact that the location of the ROI will influence the effectiveness of the extracted feature and the recognition performance. Here, we adaptively and reliably detect an ROI from the original palm image, which contains rich and stable characteristics. This step also makes the discriminative characteristics of palmprint separable from the background that contains noise and interference information. In this system, we used the hyperspectral palmprint ROI extraction method, which is based on our previous work in Ref. [36] (refer to Figure 1.4):

1. Image Enhancement: A Laplacian operator with eight neighbourhoods [37] is utilised for sharpness improvement of the original palmprint image. Afterwards, the image quality is much enhanced and will be beneficial for further preprocessing in the next steps (refer to Figure 1.4a and b). The utilised Laplacian operator is defined as follows:

  • 2. Binarisation: The Niblack [38] algorithm is a binarisation method which adaptively and locally computes the threshold of the image by performing a convolution. We first transform the enhanced palmprint image into greyscale. Then, a 2D median filter is utilised for noise reduction. In Ref. [39], it has been proved that a 2D median filter can achieve a better performance on denoising in the hyperspectral images. Lastly, we obtain the binary palm- print image using the Niblack method [38] (refer to Figure 1.4b and c).
  • 3. Palm Detection: Given the binarisation palmprint image (refer to Figure 1.4c), we initially locate the tips of the fingers (a-d) and valleys (e-h) of the palm by conducting the method in Ref. [40]. Afterwards, we detect the maximum inscribed circle (T) of the palm to find the centre of the palm (see Figure 1.4c). Therefore, the location of the maximum inscribed circle in the enhanced image can be achieved (refer to Figure 1.4d). To acquire pixels from the background, four external tangent circles of T are located as /?,, B2, Bh and #4 (see Figure 1.4c), which are on the vertical and horizontal directions. We define the radius of T as R thus, the radiuses of Bb B2, J93, and B4 are defined as 0.5 x R, 0.5 x R, 0.5 x R, and 0.3 x R, correspondingly. Here, pixels in T are randomly selected as the positive data, and pixels in Bh B2, Bh and B4 are randomly selected as the negative data. Afterwards, the positive data and the negative data are put into the SVM to segment the palm from the background (see Figure 1.4e).
  • 4. Contour Detection and ROI Extraction: Given the detected palm image (refer to Figure 1.4e), the Canny operator is utilised to achieve the boundary of the palm in the original image. Then, the boundaries named GAP, and GAP2 between the forefinger and second finger and the fourth finger
Steps of hyperspectral palmprint ROI extraction [30]

FIGURE 1.4 Steps of hyperspectral palmprint ROI extraction [30].

and little finger are obtained using the method in Ref. [41] (see Figure 1.4f), respectively. A line can then be drawn through one point in GAP! and another point in GAP2, simultaneously. Then, we can define the two key points P and A when all points in GAP and GAP2 are below the line (see Figure 1.4f). Afterwards, a coordinate system is constructed based on P| and A that the midpoint of line fl-A is defined as the origin О and a vertical line with A~A passing О is defined as the x-axis (see Figure 1.4g). At last, a sub-image with a size of 128 x 128 in the palm centre is separated from the


image using the constructed coordinate system, where OC = — A A (as seen in Figure 1.4g and h). ^

Hyperspectral Dorsal Hand Vein ROI Extraction

In the dorsal hand image, ROI indicates to the area that simply includes the vein part applied to extract feature. Dorsal hand vein images gathered through the acquisition device covers much redundant information such as a complicated background, the wrist, and the thumb. The unnecessary information can be eliminated by cropping the ROI from the collected image. The ROI not only maintains the vein structure with noise decreased but also reduces the computation cost, which can improve the recognition performance. The procedures of hyperspectral dorsal hand vein ROI extraction are presented in the following, which is adapted from our earlier study in Ref. [42] (refer to Figure 1.6):

  • 1. Pinky Knuckle Point Detection: Based on a dorsal hand in the closed state (refer to Figure 1.6a), bulges at joints of the fingers and the boundary of a dorsal hand can be taken into consideration when locating the ROI. Here, the ROI can be extracted by locating one invariant point combined with a line of the profile of the dorsal hand. To this end, the template (refer to Figure 1.5) is constructed to search the point on a pinky knuckle. Based on a correlation operation between a template and a dorsal hand image, the maximal response (see the red point denoted in Figure 1.6f) can be found as the invariant point of the pinky knuckle.
  • 2. Dorsal Hand Profile Location: The binarisation of a dorsal hand vein image was required for foreground segmenting from its background (refer to Figure 1.6b). Then, morphological opening and closing operations were applied to eliminate minor holes and remove tinny protrusions in the contour of the image (refer to Figure 1.6c). From the largest connected area (refer to Figure 1.6d), a profile of a dorsal hand (refer to Figure 1.6e) can be found by a boundary through single pixel-wise searching.
  • 3. Key Line Determination: A circle was drawn with its centre at the point of the detected pinky knuckle, where the two crossing dots between the circle and the dorsal hand profile are located (refer to Figure 1.6f). A point was found concerning a lower area of a dorsal hand, which is connected with the pinky knuckle formed a closely horizontal line. Another point was searched regarding a higher reign of a dorsal hand, which is connected with the point on the pinky knuckle produced a closely vertical line. Here,
The template to locate the pinky knuckle

FIGURE 1.5 The template to locate the pinky knuckle.

The steps of hyperspectral dorsal hand vein ROI extraction [42]

FIGURE 1.6 The steps of hyperspectral dorsal hand vein ROI extraction [42].

we chose a horizontal line or vertical line in place of an edge of an ROI (refer to Figure l .6g).

4. ROI Extraction: Finally, with the pinky knuckle point detected and one key line drawn, the other three edges of the ROI are determined (refer to Figure l.6h). Due to the insufficient vein information in margin of a dorsal hand image, the ROI is moved a few pixels to the up and right to achieve rich vascular features (refer to Figure 1.6i). The experiments showed that this method is robust and adaptive at locating the ROI precisely for hyper- spectral dorsal hand image (refer to Figure 1.6j).

< Prev   CONTENTS   Source   Next >