Self-Attention Based Approach to Iris Segmentation

This research investigates the integration of advanced self-attention techniques into deep neural network architectures, focusing on accurately segmenting the iris and pupil regions in infrared imagery. We propose modified variants of non-local block self-attention modules, tailored to enhance the self-attentive capabilities of the model while addressing the unique properties of infrared imaging data. Through these custom modifications, we achieved notable enhancements in the model's segmentation accuracy. Extensive testing was conducted using a well-curated infrared image dataset, confirming the robustness of the proposed approach. These findings open new avenues for advancements in infrared image processing, enabling diverse applications in fields ranging from biometrics to surveillance. Furthermore, this work demonstrates the potential of specialized self-attention mechanisms to significantly elevate the performance of deep learning models when applied to niche imaging challenges.