Author Affiliations
1School of Electronic & Information Engineering, Liaoning Technical University, Huludao, Liaoning 125105, China;2Fuxinlixing Technology Company Limited, Fuxin, Liaoning 123000, Chinashow less
Fig. 1. Image collection method. (a) Direct light collection; (b) light reflection collection
Fig. 2. Finger vein pretreatment process
Fig. 3. Gray scale linear interpolation schematic
Fig. 4. Finger vein effect after CLAHE algorithm processing. (a) Original image; (b) image enhancement
Fig. 5. AlexNet model structure
Fig. 6. Schematic of SPP
Fig. 7. Im-AlexNet model structure renderings
Fig. 8. Curves of JY_DB. (a) Loss curves; (b) recognition accuracy curves
Fig. 9. Curves of SD_DB. (a) Loss curves; (b) recognition accuracy curves
Layer | Number of filters | Size of kernel | Stride |
---|
Conv1 | 96 | 11×11 | 4 | Pool1 | - | 3×3 | 2 | Conv2 | 256 | 5×5 | 1 | Pool2 | - | 3×3 | 2 | Conv3 | 384 | 3×3 | 1 | Conv4 | 384 | 3×3 | 1 | Conv5 | 256 | 3×3 | 1 | Pool3 | - | 3×3 | 2 | FC layer1 | 4096 | 1×1 | 1 | FC layer2 | 4096 | 1×1 | 1 | FC layer3 | 1000 | 1×1 | 1 |
|
Table 1. AlexNet model structure parameters
Layer | Number of filters | Size of kernel | Stride |
---|
Conv1 | 96 | 5×5 | 4 | Conv2 | 96 | 1×1 | 1 | Pool1 | - | 3×3 | 2 | Conv3 | 256 | 5×5 | 1 | Conv4 | 256 | 1×1 | 1 | Pool2 | - | 3×3 | 2 | Conv5 | 384 | 3×3 | 1 | Conv6 | 384 | 1×1 | 1 | Conv7 | 384 | 3×3 | 1 | Conv8 | 384 | 1×1 | 1 | Conv9 | 256 | 3×3 | 1 | Conv10 | 256 | 1×1 | 1 | Pool3(SPP) | - | - | - | FC layer1 | 1024 | 1×1 | 1 | FC layer2 | Class | 1×1 | 1 |
|
Table 2. Im-AlexNet model structure parameters
Table 3. Finger vein datasets division
Parameter | Value |
---|
Batch | 128 | Epoch | 500 | Numbers of samples per epoch | 1280 | LR of SGD | 0.001 | Momentum of SGD | 0.9 | Training number | 64000 | Dropout | 0.25 |
|
Table 4. Experimental parameters
Network | Database | Time/min |
---|
AlexNet | JY_DB | 125 | | SD_DB | 84 | Im-AlexNet | JY_DB | 29 | | SD_DB | 25 |
|
Table 5. Comparison of training time before and after network improvement
Network | Database | Accuracy/% |
---|
AlexNet | JY_DB | 95.04 | | SD_DB | 94.36 | Im-AlexNet | JY_DB | 99.25 | | SD_DB | 98.23 |
|
Table 6. Comparison of recognition accuracy before and after network improvement
Method | Accuracy /% |
---|
SPF[24] | 87.00 | SPCF[25] | 92.71 | CLAHE+directional dilation (DD) [26] | 90.72 | CNN (ULDFV-VGG16[15]) | 92.60 | CNN (ULDFV-Xception[15]) | 93.50 | CNN (ULDFV-ResNet[15]) | 96.60 | CNN (Inception-ResNet V2[15]) | 96.70 | Dula-sliding window+location+Pseudo-elliptical transformer+2D-PCA[27] | 97.02 | Block-based average absolute deviation (AAD) features[28] | 97.76 | CNN (Proposed CNN[29]) | 97.48 | CNN (AlexNet) | 94.36 | CNN (Im-AlexNet) | 98.23 |
|
Table 7. Comparison of recognition accuracy of different image feature algorithms on SD_DB