Fig 1.
Main lines and fold characteristics of palmprint (left) and ROI (right).
Fig 2.
Illustration of the method of extracting ROI by keypoints positioning in literature.
Fig 3.
Flow structure of the proposed method.
Fig 4.
Example images of SCAUPD.
Fig 5.
Example images of the database used: (A) REST; (B) MPD; (C) IITD; (D) BMPD.
Fig 6.
YOLOv5-lite network structure.
Fig 7.
YOLOv5-lite palm initial localization output.
Fig 8.
UNet network structure.
Fig 9.
Proposed network structure.
Fig 10.
Example of residual downsampling, input size (h, w, c), output size (h/2, w/2, c).
Fig 11.
Depthwise separable convolution.
Fig 12.
Keypoints in the original image (left) and keypoints visualized under the heatmap (right).
Fig 13.
Schematic diagram of ROI extraction algorithm.
Fig 14.
Example of data annotation.
Table 1.
Statistics of the dataset used in the experiment.
Fig 15.
Training and validation loss curves for YOLOv5-lite.
Fig 16.
Training and validation accuracy curves for YOLOv5-lite.
Table 2.
Comparison of experimental results between YOLOv5-lite and other lightweight detection networks.
Fig 17.
Training and testing loss function curves for the improved UNet model.
Fig 18.
Training and testing accuracy curves for the improved UNet model.
Fig 19.
Threshold = 2 model output image (left) and normal label model output image (right).
Table 3.
Model training accuracy and test accuracy under different thresholds.
Fig 20.
The implementation effect of the model on different datasets when Threshold = 1.5.
(A) is BMPD, (B) is IITD, (C) is MPD, (D) is REST, and (E) is SCAUPD.
Fig 21.
Convergence curves of model accuracy under different loss functions.
Table 4.
Comparison of test accuracies for the model under different loss functions.
Fig 22.
Convergence curve of accuracy for the residual block ablation experiment.
Table 5.
Comparison of model size, average detection time, and testing accuracy using different modules.
Fig 23.
Diagram of changes in testing accuracy between the proposed network and other networks.
Table 6.
Comparison between the proposed network and other networks.
Table 7.
Open set testing of the proposed method compared to other advanced methods.
Fig 24.
The performance of the proposed method on the RPG1K dataset.
Fig 25.
The performance of the proposed method in the example of PKLNet [26] localization failure.
Table 8.
The average detection time and model size of the proposed method compared to other advanced methods.