📝 Representative Publications

Daoxiang Zhou*, Feixue Shao, Jun Zhang, Yunfei Zhang, Shu Feng.
Multi-scale orthogonal Gabor filters based ConvNets for illumination robust single sample face recognition. Engineering Applications of Artificial Intelligence (SCI, Q1-TOP). | CODE

• We propose a novel PFCNN method with orthogonal Gabor filters.
• We orthogonalize the Gabor filters to enhance the noise robustness and discriminative power.
• We utilize the 5 scales of Gabor filters separately, embedding multi-scale information under a single filter size.

Jun Zhang, Jia Shen, Hui Wang, Jiahua Liu, Yixing Gao, Shu Feng, Dongxi Li, Daoxiang Zhou*.
Heterogeneous Feature Mutual-Calibration Assisted Online Distillation for Efficient Face Anti-Spoofing. ICASSP (CCF-B). | CODE

• We propose a novel online distillation framework, enabling bidirectional feature transfer and collaborative training between teacher and student.
• We design a cascaded local-global feature extractor that enhances the representation capacity of teacher and guides the student to learn both texture details and global information of face image.
• We introduce a feature mutual calibration distillation loss to deal with intermediate feature mismatches between teacher and student.

Yunfei Zhang, Jun Zhang, Tonglu Li, Feixue Shao, Xuetao Ma, Yongfei Wu, Shu Feng, Daoxiang Zhou*.
ACAN: A Plug-and-Play Adaptive Center-Aligned Network for Unsupervised Domain Adaptation. Engineering Applications of Artificial Intelligence (SCI, Q1-TOP). | CODE

• It is a plug-and-play module and can be easily incorporated into any domain alignment methods.
• Angular margin loss is called to enhance the discrimination power for classifier.
• Entropy regularization is exploited to highlight the probability of potential related class.

Jun Zhang, Yunfei Zhang, Feixue Shao, Xuetao Ma, Shu Feng, Yongfei Wu, Daoxiang Zhou*.
Efficient face anti-spoofing via head-aware transformer based knowledge distillation with 5 MB model parameters. Applied Soft Computing (SCI, Q2-TOP). | CODE

• Feature- and logits-level distillation are combined to transfer the rich knowledge of teacher to student.
• A head-aware strategy is proposed to deal with the dimension mismatching issue of middle encoder layers between teacher and student networks.
• Our student network is extremely lightweight with only 5 MB parameters.

Daoxiang Zhou*, Shu Feng.
M3SPCANet: A simple and effective ConvNets with unsupervised predefined filters for face recognition. Engineering Applications of Artificial Intelligence (SCI, Q1-TOP). | CODE

• We present a novel lightweight CNN alternative with predefined PCA filters to learn face feature representation.
• We learn multi-scale features that are complement with each other with stacking strategy.
• We leverage spatial pyramid SOP for high-order correlation information between each pair of channel features.

Zhigang Yang, Daoxiang Zhou, Ying Yang, Jiapeng Zhang, Zehua Chen*.
Road extraction from satellite imagery by road context and full-stage feature. IEEE Geoscience and Remote Sensing Letters (ESI，高被引论文). | CODE

• We propose the MSCE module to capture the long-range dependencies, which can efficiently capture road context over long distances.
• Multiple FSFF in skip connection that integrates features in different stages are constructed.
• We devise a coordinate dual attention mechanism in FSFF to strengthen the channel and spatial features of roads.