Lightweight Image Pre-processing Model for Improving Segmentation of Infrared Human Eye Images

This study focuses on improving iris and pupil segmentation in infrared images, a crucial task for gaze-tracking systems. We propose a lightweight image preprocessing approach, which ensures controlled data transformation, effectively eliminating parasitic reflections, noise, and contrast inconsistencies without introducing unwanted artifacts. Our approach employs analytically defined transformations, incorporating local and global filtering techniques to enhance the quality of input images while preserving physiologically relevant features. We compared deep learning-based segmentation models with and without preprocessing, demonstrating that the proposed approach significantly improves performance. Experimental results show notable gains in the mIoU metric, confirming increased robustness to illumination and image quality variations.