Fig. 1. Schemetic diagram of the multispectral palmprint imaging device
Fig. 2. Multispectral palmprint ROI extraction process
(a): Near-infrared palmprint; (b): Binarization; (c): Maximum incision circle; (d): Near-infrared palmprint ROI
Fig. 3. Multispectral palmprint ROI extraction process
Fig. 4. Mapping of local maxima and local minima
(a): Source signal; (b): Local maxima map; (c): Local minima map
Fig. 5. Decompositions of a palmprint image using FABEMD
(a): Green ROI; (b): BIMF1; (c): BIMF2; (d): BIMF3; (e): BIMF4; (f): The residue; (g): The sum of K BIMF
Fig. 6. Near-infrared palmprint image processing
(a): NIR ROI; (b): SI; (c): ${{\tilde{S}}_{I}}$(Bezier interpolation); (d): S'I(feature refinement)
Fig. 7. Multispectral palmprint fusion image
(a): Fused images without feature compression;(b): Fused images with feature compression
Fig. 8. Multispectral palmprinu fusion process
(a): SR (red in FABEMD); (b): SG; (c): SB; (d): S″I(NIR with feature compression); (e): Fusion process
Fig. 9. Network architecture of improved residual channel attention network (IRCANet)
Fig. 10. Original ResBlock and ResStageBlock
Fig. 11. Improved residual block with channel attention
Fig. 12. Multispectral palmprint samples in the PolyU database
Fig. 13. Accuracies of different K values in FABEMD
Fig. 14. Accuracies of different fusion strategies for multispectral (R, G, B, N) palmprint images
Fig. 15. Optimal results of different residual structures for multispectral (R, G, B, N) palmprint images
Fig. 16. Recognition effect of different mainstream networks for multispectral palmprint fusion images
(a): Recognition trends; (b): Accuracy contrast
学习率 | 批大小 | 迭代次数 | 平均精准度/% |
---|
0.001 | 4 | 100 | 97.88 | 0.001 | 4 | 200 | 98.93 | 0.001 | 6 | 100 | 99.67 | 0.001 | 6 | 200 | 99.59 | 0.000 1 | 8 | 100 | 99.62 |
|
Table 1. Average accuracies of different IRCANet parameters
融合光谱带 | 精准度/% |
---|
直接加权 融合 | 红外背景 重建 | 红外背景重建 及特征细化 |
---|
Red, NIR | 92.04 | 95.08 | 99.49 | Green, NIR | 90.63 | 93.35 | 99.22 | Blue, NIR | 91.72 | 94.41 | 99.37 | Red, Green, NIR | 91.12 | 94.92 | 99.05 | Blue, Red, NIR | 90.76 | 95.01 | 99.45 | Blue, Green, NIR | 90.68 | 93.77 | 99.19 | Blue, Green, Blue, NIR | 90.92 | 95.12 | 99.67 |
|
Table 2. Accuracies of multispectral combinations under different fusion strategies
融合光谱带 | 精准度率/% |
---|
传统残差 结构 | 分阶段残 差结构 | 结合注意力 机制分阶段 残差结构 |
---|
Red, Blue | 96.37 | 91.23 | 99.25 | Green, Red | 95.57 | 90.55 | 99.41 | Blue, Green | 93.10 | 89.58 | 99.51 | Red, NIR | 96.93 | 92.66 | 99.09 | Green, NIR | 95.03 | 91.11 | 99.22 | Blue, NIR | 95.70 | 90.99 | 99.37 | Red, Green, Blue | 96.79 | 92.13 | 99.10 | Red, Green, NIR | 94.71 | 90.60 | 99.05 | Blue, Red, NIR | 95.36 | 91.51 | 99.55 | Blue, Green, NIR | 96.85 | 91.74 | 99.49 | Blue, Green, Blue, NIR | 94.03 | 90.80 | 99.67 |
|
Table 3. Accuracies of multispectral combinations under different residual structures
算法 | Accuracy/% |
---|
Image-level fusion by DWT[18] | 99.03 | Matching score-level fusion[10] | 99.43 | QPCA+QDWT[19] | 98.83 | Deep Scattering Network[20] | 99.40 | Joint Pixel and Feature Alignment[21] | 99.16 | Inception-ResNet[22] | 98.74 | 本算法 | 99.67 |
|
Table 4. Comparison of multispectral palmprint recognition algorithms
算法步骤 | 平均运行时间/s |
---|
FABEMD | 0.175×4 | 近红外背景重建及特征细化 | 0.186 | 图像融合 | 0.002 1 | 分类识别 | 0.001 5 | 总的识别时间 | 0.889 6 |
|
Table 5. Time cost of the proposed method