Unleashing the Power of Dense Uncertainty Embeddings for More Efficient and Accurate Iris Recognition

Pixelwise dense representations are more prevalent in the field of iris recognition, also known as iris templates or IrisCodes. Almost all previous works of this kind are deterministic. To be specific, pixel-level representations are exclusively derived from certain point-by-point modeling, including filter responses, phase correlations, and ordinal relations. Moreover, the binary mask indicating valid iris regions is solely determined by a fixed threshold or the output of standalone segmentation and localization algorithms. Uncertainty in acquisition factors in the process of iris imagery formation is not considered. In this paper, we propose a simple yet effective plug-and-play building block termed dual dense uncertainty embedding (D2UE), which can be seamlessly incorporated into deep learning (DL) frameworks that extract dense representations for iris recognition. D2UE has two pathways wherein both take dense feature maps of the backbone network as input. One pathway of D2UE predicts a variance-scaling map (VSM) and then applies it to an adaptive threshold-masking operation on the iris image. The dynamic threshold for each pixel in this manner is dependent on not only the intensity distribution of the iris image but also each pixel’s low-level uncertainty. The other pathway of D2UE adopts an over-parameterization technique and extracts uncertainty-embedded dense representations (UEDRs) by modeling each pixel’s contextual uncertainty. Extensive experiments on several iris datasets demonstrate that recognition performance under both within-database and cross-database settings can be significantly improved by incorporating D2UE into the baseline method. By integrating D2UE into various deep learning frameworks and evaluating their performance across multiple datasets, the results demonstrate that D2UE can be seamlessly incorporated into diverse architectures and can significantly enhance their recognition capabilities. D2UE only incurs slight computational overhead while surpassing a few SOTA methods with a large backbone network and much more training budget.