Conversion of Iris Image from Off-Angle to Frontal Using Hybrid Attention and Multi-Scale-Discriminator GAN

Abstract

Iris recognition has extensive applications in border control, healthcare, mobile authentication, and security systems. In real-world situations, the captured iris images often suffer from distortions, occlusions (by eyelids or blinking), and off-angle captures due to acquisition angle and the eye’s 3D structure, leading to information loss. These off-angle images significantly reduce recognition accuracy, making frontal view reconstruction a critical preprocessing step. Although GANs have shown promise in this area, they struggle to handle extreme gaze angles (greater than 30°), occlusions caused by eyelids or blinking, and distortions arising from the 3D structure of the eye. To address these challenges, this study proposed a GAN-based approach that integrates attention mechanisms and a multi-scale discriminator to convert off-angle iris images to high-quality frontal views. The generator is specifically based on a U-Net architecture improved by introducing channel attention in the encoder and self-attention in the decoder. Meanwhile, the discriminator utilizes a multi-scale approach to capture local and global features more effectively. The proposed method is tested on the MOBIUS and UBIRIS.v2 datasets and compared with baseline models. Quantitative measurement of the quality of the converted iris images was done using metrics such as Peak Signal-to-Noise Ratio, Structural Similarity Index Measure , and Hamming Distance. Our proposed model significantly outperforms the baseline Pix2Pix-GAN across these metrics. For PSNR, it improves from 28.00 to 30.34 (-30°), 28.01 to 30.31 (30°), 22.97 to 27.99 (-45°), and 23.97 to 28.98 (45°). SSIM improves from 0.69 to 0.79, 0.70 to 0.80, 0.61 to 0.76, and 0.60 to 0.75, respectively. Hamming Distance is reduced from 0.47 to 0.34, 0.44 to 0.36, 0.49 to 0.42, and 0.50 to 0.44. Additionally, our model achieves overall improvements with PSNR rising from 18.75 to 28.94, SSIM from 0.61 to 0.82, and achieves a lower Hamming Distance of 0.44 from 0.97 for converting occluded iris images caused by eyelids, partially visible iris images, and distorted iris images caused by its 3D structure of the eye or due to a large angle to frontal iris images. These results indicate that our model is more effective at handling largeangle up to 45°, partially occluded, and structurally distorted iris images, making it a valuable advancement for iris recognition and surveillance systems that require accurate frontal iris representations.