Uncertainty-Aware Bilateral Transformer for Accurate and Reliable Iris Segmentation IF: 6.8

Jianze Wei; Xingyu Gao; Yunlong Wang; Ran He; Zhenan Sun

• UTIris, a new iris dataset with high acquisition and annoation uncertainties, is built, and we provide rich annotations including two levels (rough- and precise-levels) of binary maps for iris masks, the circles of iris inner and outer boundaries, and curves of upper and lower eyelids to support future research on it.
• A bilateral self-attention module is proposed to capture visual and spatial relationships explicitly. Based on the module, a new backbone named BiTrans is designed to reduce the negative influence of the acquisition uncertainty in iris segmentation.

Multi-Faceted Knowledge-Driven Graph Neural Network for Iris Segmentation IF: 6.8

Jianze Wei; Yunlong Wang; Xingyu Gao; Ran He; Zhenan Sun

• A graph-based method with U-net structure for iris segmentation.
• Summary: This paper utilizes the multi-faceted knowledge of the image, including including visual similarity, positional correlation, and semantic consistency, to construct self-adaptive relationships for accurate iris segmentation.

Contextual Measures for Iris Recognition IF: 6.8

Jianze Wei, Yunlong Wang, Huaibo Huang, Ran He, Zhenan Sun, Xingyu Gao

• The first Transformer model for iris recognition.
• Summary: the paper integrates the advantages of visual Transformer and CNN, and proposes contextual measures (CM). The proposed CM regards each iris region as a potential microstructure and models the correlations between them for iris recognition.

Towards More Discriminative and Robust Iris Recognition by Learning Uncertain Factors IF: 6.8

Jianze Wei, Huaibo Huang, Yunlong Wang, Ran He, Zhenan Sun

Project

• Summary: the paper represents an iris image using a probabilistic distribution rather than a deterministic point (binary template or feature vector) that is widely adopted in iris recognition methods.
• Extension: the proposed representation augments input data in the feature level, and it is employed in Contrastive Uncertainty Learning for Iris Recognition with Insufficient Labeled Samples

Cross-Spectral Iris Recognition by Learning Device-Specific Band IF:8.4

Jianze Wei, Yunlong Wang, Yi Li, Ran He, Zhenan Sun,

Project

• Summary: the paper proposes a Gabor Trident Network (GTN) to narrow the distribution gap between near-infrared (NIR) and visible (VIS) images. GTN first utilizes the Gabor function’s priors to perceive iris textures under different spectra, and then codes the device-specific band as the residual component to assist the generation of spectral-invariant features.

Cross-sensor Iris Recognition Using Adversarial Strategy and Sensor-specific Information EI, Oral

Jianze Wei, Yunlong Wang, Xiang Wu, Zhaofeng He, Ran He, Zhenan Sun

• Summary: the paper propose Cross-sensor iris network (CSIN) by applying the adversarial strategy and weakening interference of sensor-specific information.

Contrastive Uncertainty Learning for Iris Recognition with Insufficient Labeled Samples CCF-C, Oral

Jianze Wei, Ran He, Zhenan Sun

Project

• The first work for both unsupervised and semi-supervised iris recognition method.
• Summary: the paper explores the uncertain acquisition factors and adopts a probabilistic embedding to represent the iris image, then it utilizes this probabilistic representation to generate virtual positive and negative pairs.

Learning Discriminative Geodesic Flow Kernel for Unsupervised Domain Adaptation CCF-B, Oral

Jianze Wei, Jian Liang, Ran He, Jinfeng Yang

• Summary: the paper extends the classic geodesic flow kernel method by leveraging the pseudo labels during the training process to learn a discriminative geodesic flow kernel for unsupervised domain adaptation.

Diagnostic Classification for Human Autism and Obsessive-Compulsive Disorder Based on Machine Learning From a Primate Genetic Model IF:17.7

Yafeng Zhan, Jianze Wei (co-first), Jian Liang, Xiu Xu, Ran He, Trevor W Robbins, Zheng Wang

• Summary: the paper develops a new group lasso method to search the share egg featurs between humans and monkeys for ASD diagnosis. Specifically, our method groups all connection features into 94 groups according to their connected brain regions. Then, a group lasso is employed and elaborately set to find the effective features. Finally, a classifier model trained on monkey data is leveraged to predict the diagnostic results of ASD.
• Academic Impact: ESI Highly Cited Paper.

Deep Mutual Distillation for Unsupervised Domain Adaptation Person Re-Identification IF: 8.4

Xingyu Gao; Zhenyu Chen; Jianze Wei; Rubo Wang; Zhijun Zhao

• deep mutual distillation (DMD) to generate reliable pseudo-labels for unsupervised domain adaptation person re-identification.
• Summary: the paper proposes a deep mutual distillation (DMD) to generate reliable pseudo-labels for UDA person re-ID. The proposed DMD applies two parallel branches for feature extraction, and each branch serves as the teacher of the other to generate pseudolabels for its training.