Toward Accurate and Reliable Iris Segmentation Using Uncertainty Learning

TL;DR

This paper introduces a novel iris segmentation method combining a Bilateral Transformer with uncertainty learning, significantly improving accuracy and reliability while reducing computational cost compared to state-of-the-art models.

Contribution

The paper proposes a Bilateral Transformer architecture with a bilateral self-attention module and an uncertainty learning scheme for more accurate and reliable iris segmentation.

Findings

Achieves better segmentation performance than SOTA with 20% FLOPs

Effectively estimates segmentation uncertainty reflecting prediction reliability

Improves spatial perception and feature fusion in iris segmentation

Abstract

Iris segmentation is a deterministic part of the iris recognition system. Unreliable segmentation of iris regions especially the limbic area is still the bottleneck problem, which impedes more accurate recognition. To make further efforts on accurate and reliable iris segmentation, we propose a bilateral self-attention module and design Bilateral Transformer (BiTrans) with hierarchical architecture by exploring spatial and visual relationships. The bilateral self-attention module adopts a spatial branch to capture spatial contextual information without resolution reduction and a visual branch with a large receptive field to extract the visual contextual features. BiTrans actively applies convolutional projections and cross-attention to improve spatial perception and hierarchical feature fusion. Besides, Iris Segmentation Uncertainty Learning is developed to learn the uncertainty map according to prediction discrepancy. With the estimated uncertainty, a weighting scheme and a regularization term are designed to reduce predictive uncertainty. More importantly, the uncertainty estimate reflects the reliability of the segmentation predictions. Experimental results on three publicly available databases demonstrate that the proposed approach achieves better segmentation performance using 20% FLOPs of the SOTA IrisParseNet.