Challenges in Off-angle to Frontal Iris Image Conversion using Pix2Pix Generative Adversarial Networks

Person identification using biometrics has become a safer and trustworthy mechanism with the advancement of technology. Among all biometric identification methods, iris recognition has achieved very low false acceptance rates due to its complex and unique patterns. The low acceptance rates apply only to frontal iris images. Capturing frontal iris images is not always possible, especially in uncontrolled environments, where most of the iris images captured tend to be non-ideal, such as off-angle images. Off-angle iris images suffer from several issues, including corneal refraction, limbus occlusion, the effect of gaze angle, and depth of field blur. These effects distort the iris patterns, causing the similarity scores between the same individual to widen and scores between different individuals to become closer. This also causes false acceptance rates to increase, as it increases the chances of misclassification. This highlights the need for improving the performance of off-angle iris recognition. By leveraging the low false-acceptance rates of the frontal iris images, we build generated frontal version of the iris images using off-angle iris images and achieved better performance compared with the perspective transformation. We built a modified version of the Pix2Pix GAN to achieve the frontal projection of off-angle iris images. Instead of using a Mean Squared loss function in the Pix2Pix GAN, we use a combination of Mean Squared loss function, Matrix Multiplication loss, and SSIM loss function to generate sharper images that can capture the textural information of the original image better.