UniRotNet: Mask-only canonical alignment for contactless palm ROI localization

  • Unique Paper ID: 210748
  • Volume: 13
  • Issue: 6
  • PageNo: 2233-2252
  • Abstract:
  • Reliable palm region extraction is a critical prerequisite for contactless palm biometrics, where pose variation and background clutter can severely degrade matching accuracy. Many high-performing palmprint and palm-vein recognition methods particularly those based on deep convolutional features and texture coding rely on geometrically normalized inputs and therefore depend on a stable preprocessing front-end. This paper presents a mask-guided three-stage pipeline tailored to this scenario: (i) a U-Net segments the hand from the background, (ii) a lightweight network (UniRotNet) enforces a canonical upright pose by estimating in-plane rotation from the predicted binary mask alone, and (iii) lightweight CNN regressors localize a square palm region of interest (ROI) from the aligned mask. By operating on binary masks in Stages II and III, the design reduces the dependence of alignment and ROI localization on raw appearance. Experiments on the CASIA Multi-Spectral Palmprint database, using subject-disjoint splits over three random seeds, yield high segmentation accuracy (IoU = 0.9911 ± 0.0011), precise mask-only rotation correction (MAE = 3.004 ± 0.567°), and accurate ROI localization on aligned masks (mIoU = 0.909 ± 0.002) with negligible outside-of-hand leakage. Compared with an end-to-end regressor trained on raw images with axis-aligned box targets (Image-PalmROI), the mask-guided regressors give tighter ROIs (mIoU = 0.905–0.909 vs 0.641) and, on images aligned with the recorded hand angle, lower downstream matching error (4.11% vs 9.66% EER with a fine-tuned ResNet-18 matcher), indicating that mask-guided canonical alignment provides a stable front-end for contactless palm biometrics on this dataset.

Copyright & License

Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

BibTeX

@article{210748,
        author = {Amani Abbas Almaktari and Mohammed Bokash and Vikas T. Humbe},
        title = {UniRotNet: Mask-only canonical alignment for contactless palm ROI localization},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {6},
        pages = {2233-2252},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=210748},
        abstract = {Reliable palm region extraction is a critical prerequisite for contactless palm biometrics, where pose variation and background clutter can severely degrade matching accuracy. Many high-performing palmprint and palm-vein recognition methods particularly those based on deep convolutional features and texture coding rely on geometrically normalized inputs and therefore depend on a stable preprocessing front-end. This paper presents a mask-guided three-stage pipeline tailored to this scenario: (i) a U-Net segments the hand from the background, (ii) a lightweight network (UniRotNet) enforces a canonical upright pose by estimating in-plane rotation from the predicted binary mask alone, and (iii) lightweight CNN regressors localize a square palm region of interest (ROI) from the aligned mask. By operating on binary masks in Stages II and III, the design reduces the dependence of alignment and ROI localization on raw appearance. Experiments on the CASIA Multi-Spectral Palmprint database, using subject-disjoint splits over three random seeds, yield high segmentation accuracy (IoU = 0.9911 ± 0.0011), precise mask-only rotation correction (MAE = 3.004 ± 0.567°), and accurate ROI localization on aligned masks (mIoU = 0.909 ± 0.002) with negligible outside-of-hand leakage. Compared with an end-to-end regressor trained on raw images with axis-aligned box targets (Image-PalmROI), the mask-guided regressors give tighter ROIs (mIoU = 0.905–0.909 vs 0.641) and, on images aligned with the recorded hand angle, lower downstream matching error (4.11% vs 9.66% EER with a fine-tuned ResNet-18 matcher), indicating that mask-guided canonical alignment provides a stable front-end for contactless palm biometrics on this dataset.},
        keywords = {Canonical alignment, Hand segmentation, Palm biometrics, ROI localization, Rotation correction.},
        month = {November},
        }

Cite This Article

Almaktari, A. A., & Bokash, M., & Humbe, V. T. (2026). UniRotNet: Mask-only canonical alignment for contactless palm ROI localization. International Journal of Innovative Research in Technology (IJIRT). https://doi.org/doi.org/10.64643/IJIRT210748

Related Articles