Lite-HDNet: A lightweight domain-adaptive segmentation framework for improved finger vein pattern extraction

• Published in: IEEE access Vol. 12; p. 1
• Main Authors: Li, Yingxin, Chen, Yucong, Zeng, Junying, Qin, Chuanbo, Zhang, Wenguang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.01.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adaptation models; Data mining; Datasets; Deep learning; Domain Adaptation; Feature extraction; Finger Vein Extraction; Fingers; Heterogeneity; Image quality; Image segmentation; Knowledge management; Knowledge transfer; Lightweight; Real time; Real-time systems; Representations; Training; Trunk networks; Veins
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3382197

Recent times have witnessed significant progress in deep learning-based finger vein pattern extraction methods, but two unavoidable issues still remain to be addressed. One is that the model trained on a single finger vein dataset shows poor generalizability, and the model performance is limited by the image quality of the single dataset; the other is that it is hard for the deep model to extract real-time finger vein patterns because of its large number of parameters and poor real-time performance. To address the aforementioned issues, we propose a novel lightweight domain-adaptive segmentation framework (Lite-HDNet) that learns a generic representation of different domains to improve the extraction of finger vein patterns. We propose a multi-domain feature knowledge transfer strategy and a domain migration loss converter to enable the trunk network to learn the robust representations of different finger vein datasets as well as to compensate for the heterogeneity between them. In the proposed framework, two lightweight segmentation networks are designed as the trunk branch and the auxiliary branch to achieve real-time extraction of finger vein patterns. Our approach has been extensively tested on four finger vein datasets available to the public, and the results show that our Lite-HDNet not only improves segmentation performance on all datasets but also effectively reduces heterogeneity between different domains. In addition, we also validated the real-time performance of Lite-HDNet on NVIDIA embedded terminals, proving the outperformance of our approach compared with previous lightweight segmentation networks.