IRIS recognition systems are crucial for national ID programs, yet face performance issues due to diverse IRIS sensors and environmental variations. This paper presents a domain adaptation framework and a Markov random fields-based algorithm to enhance cross-domain IRIS recognition. Leveraging naive Bayes nearest neighbor classification and real-valued feature representation, the framework learns domain knowledge. By estimating visible iris patterns from synthesized near-infrared images, our approach outperforms existing methods in cross-spectral iris recognition. Furthermore, we propose and assess a bi-spectral IRIS recognition system capable of simultaneously capturing visible and near-infrared images with pixel-to-pixel correspondences.