An Ultra-lightweight Real-time Segmentation Network of Finger Vein Textures

Junying ZENG; Yucong CHEN; Xihua LIN; Chuanbo QIN; Yinbo WANG; Jingming ZHU; Lianfang TIAN; Yikui ZHAI; Junying GAN

doi:10.3788/gzxb20225102.0210009

Journals >Acta Photonica Sinica >Volume 51 >Issue 2 >Page 0210009 > Article

Acta Photonica Sinica
Vol. 51, Issue 2, 0210009 (2022)

An Ultra-lightweight Real-time Segmentation Network of Finger Vein Textures

Junying ZENG¹, Yucong CHEN¹, Xihua LIN¹, Chuanbo QIN^1,*..., Yinbo WANG¹, Jingming ZHU¹, Lianfang TIAN², Yikui ZHAI¹ and Junying GAN¹|Show fewer author(s)

Author Affiliations

¹Faculty of Intelligent Manufacturing，Wuyi University，Jiangmen，Guangdong 529020，China

²School of Automation Science and Engineering，South China University of Technology，Guangzhou 510640，China

show less

DOI: 10.3788/gzxb20225102.0210009 Cite this Article

Junying ZENG, Yucong CHEN, Xihua LIN, Chuanbo QIN, Yinbo WANG, Jingming ZHU, Lianfang TIAN, Yikui ZHAI, Junying GAN. An Ultra-lightweight Real-time Segmentation Network of Finger Vein Textures[J]. Acta Photonica Sinica, 2022, 51(2): 0210009 Copy Citation Text

show less

Fig. 1. The overall structure of the SGUnetV3

Download full size | View in the Article

Fig. 2. SGUnetV1 basic block structure diagram

Download full size | View in the Article

Fig. 3. The basic module structure after joining the Cheap operation

Download full size | View in the Article

Fig. 4. The basic block structure of SGUnetV3 network

Download full size | View in the Article

Fig. 5. Three options for ECA module placement

Download full size | View in the Article

Fig. 6. The effect of finger vein segmentation

Download full size | View in the Article

Fig. 7. Segmentation visualization of each patch of finger vein

Download full size | View in the Article

Fig. 8. The actual segmentation effect diagram of SGUnet and each lightweight network on the SDU-FV dataset

Download full size | View in the Article

Fig. 9. The actual segmentation effect diagram of SGUnet and each lightweight network on the MMCBNU_6000 dataset

Download full size | View in the Article

Fig. 10. Comparison of important indicators with classic lightweight networks

Download full size | View in the Article

Network	Params	Mult-Adds	Dice	AUC	Accuracy	Specificity	Precision
Unet	13.39M	1.928G	0.444 6	0.843 4	91.16%	96.47%	53.79%
MobileV2+Unet	5.289M	171.226M	0.502 5	0.855 4	91.31%	95.34%	53.68%
SGUnet	516.014k	39.494M	0.503 0	0.898 2	91.89%	95.95%	57.27%

Table 1. Improved network SGUnet and basic Unet，MobileV2+Unet performance comparison table

View in the Article

Network	Dice	AUC	Accuracy	Specificity	Precision
SGUnet	0.503 0	0.898 2	91.89%	95.95%	57.27%
SGUnet+SE	0.496 1	0.892 6	91.95%	96.32%	58.27%
SGUnet+CA-32	0.496 8	0.886 7	91.89%	96.25%	57.79%
SGUnet+CA-16	0.500 8	0.891 7	91.93%	96.18%	57.87%
SGUnet+CA-8	0.501 5	0.886 9	92.00%	96.30%	58.55%
SGUnet+ECA-3	0.497 8	0.891 3	91.80%	96.11%	58.37%
SGUnet+ECA-5	0.503 8	0.898 2	92.10%	96.34%	59.04%

Table 2. Performance comparison table of ECA module，classic SE module and CA attention module added on the basis of SGUnet

View in the Article

Network	Dice	AUC	Accuracy	Specificity	Precision
a	0.501 6	0.898 2	92.10%	96.34%	59.04%
b	0.501 2	0.896 4	92.03%	96.18%	58.54%
c	0.500 4	0.894 5	91.97%	96.26%	58.32%

Table 3. Comparison of network performance of three different schemes a，b，and c

View in the Article

Network	Params	Mult-Adds	Dice	AUC	Accuracy	Specificity	Precision
Unet	13.39M	1.928G	0.444 6	0.843 4	91.16%	96.47%	53.79%
MobileV1+Unet	3.932M	481.35M	0.498 9	0.855 4	91.31%	95.34%	53.68%
MobileV2+Unet	5.289M	171.226M	0.502 5	0.884 6	91.54%	95.87%	56.68%
SGUnetV1	516.054k	39.504M	0.503 8	0.898 2	92.10%	96.34%	59.04%
SGUnetV2	416.752k	26.575M	0.499 2	0.898 9	91.73%	96.12%	56.14%

Table 4. Comparison of parameters between SGUnetV2 and other networks

View in the Article

Network	Params	Mult-Adds	Dice	AUC	Accuracy	Specificity	Precision
R2UNet	48.92M	-----	-----	0.902 9	91.87%	98.21%	62.18%
DUNet	26.73M	-----	-----	0.913 3	91.99%	97.26%	64.20%
Unet	13.39M	1.928G	0.444 6	0.843 4	91.17%	96.48%	53.79%
SGUnetV1	516.054k	39.504M	0.503 8	0.898 2	92.10%	96.34%	59.04%
SGUnetV2	416.752k	26.575M	0.499 2	0.898 9	91.73%	96.12%	56.14%
SGUnetV3	145.25k	10.453M	0.497 3	0.899 2	91.60%	95.81%	55.34%

Table 5. Experimental results of SGUnet series network and large-scale network on SDU-FV data set

View in the Article

Network	Params	Mult-Adds	Dice	AUC	Accuracy	Specificity	Precision
R2UNet	48.92M	-----	-----	0.905 8	92.94%	97.22%	54.68%
DUNet	26.73M	-----	-----	0.912 5	93.30%	97.89%	58.82%
Unet	13.39M	1.928G	0.437 2	0.847 4	91.03%	95.70%	49.49%
SGUnetV1	516.054k	39.504M	0.538 4	0.934 4	94.11%	96.79%	60.44%
SGUnetV2	416.752k	26.575M	0.527 9	0.935 4	93.75%	96.31%	57.55%
SGUnetV3	145.25k	10.453M	0.520 2	0.933 3	93.68%	96.36%	57.23%

Table 6. Experimental results of SGUnet series network and large-scale network on MMCBNU_6000 data set

View in the Article

Network	Params	FLops	Mult-Adds	Dice	AUC	Accuracy	Specificity	Precision
Unet	13.39M	1.95G	1.928G	0.444 6	0.843 4	91.17%	96.48%	53.79%
Squeeze_Unet	2.893M	296.05M	287.61M	0.501 7	0.863 0	91.02%	94.72%	51.90%
Mobile_Unet	3.932M	498.13M	481.35M	0.502 5	0.855 4	91.31%	95.34%	53.68%
Ghost_Unet	6.783M	130.46M	128.97M	0.485 3	0.886 4	91.84%	96.75%	58.47%
Shuffle_Unet	516K	68.27M	57.97M	0.511 6	0.885 9	91.48%	95.28%	54.49%
SGUnetV1	516.054k	42.97M	39.504M	0.503 8	0.898 2	92.10%	96.34%	59.04%
SGUnetV2	416.752k	29.26M	26.575M	0.499 2	0.898 9	91.73%	96.12%	56.14%
SGUnetV3	145.25k	13.13M	10.453M	0.497 3	0.899 2	91.60%	95.81%	55.34%

Table 7. Experimental data of SGUnet series network and other lightweight networks on SDU-FV dataset

View in the Article

Network	Params	FLops	Mult-Adds	Dice	AUC	Accuracy	Specificity	Precision
Unet	13.39M	1.95G	1.928G	0.474 1	0.883 4	92.42%	95.80%	49.24%
Squeeze_Unet	2.893M	296.05M	287.61M	0.510 5	0.921 6	93.08%	95.34%	52.60%
Mobile_Unet	3.932M	498.13M	481.35M	0.500 3	0.905 4	92.06%	94.52%	53.37%
Ghost_Unet	6.783M	130.46M	128.97M	0.511 0	0.924 3	93.01%	96.43%	58.38%
Shuffle_Unet	516K	68.27M	57.97M	0.479 2	0.906 6	91.80%	94.15%	46.33%
SGUnetV1	516.054k	42.97M	39.504M	0.538 4	0.934 4	94.11%	96.79%	60.44%
SGUnetV2	416.752k	29.26M	26.575M	0.527 9	0.935 4	93.75%	96.31%	57.55%
SGUnetV3	145.25k	13.13M	10.453M	0.520 2	0.933 3	93.68%	96.36%	57.23%

Table 8. Experimental data of SGUnet series network and other lightweight networks on MMCBNU_6000 dataset

View in the Article

Network	Time/s
Network	NANO	TX2	NX	AGX
Squeeze_Unet	5.146	0.686	0.663	0.388
Mobile_Unet	5.699	0.519	0.505	0.288
Ghost_Unet	2.719	0.684	0.722	0.358
Shuffle_Unet	3.208	0.766	0.654	0.426
SGUnetV1	2.735	0.455	0.386	0.283
SGUnetV2	2.789	0.523	0.427	0.302
SGUnetV3	2.827	0.569	0.458	0.306

Table 9. The running time of SGUnet series and other lightweight networks to process a single SDU-FV data set image on the NVIDIA embedded platform

View in the Article

Network	Time/s
Network	NANO	TX2	NX	AGX
Squeeze_Unet	5.241	0.679	0.699	0.357
Mobile_Unet	5.435	0.524	0.737	0.284
Ghost_Unet	2.791	0.612	0.734	0.346
Shuffle_Unet	3.221	0.766	0.652	0.418
SGUnetV1	2.779	0.457	0.405	0.270
SGUnetV2	2.873	0.525	0.418	0.290
SGUnetV3	2.928	0.570	0.467	0.291

Table 10. The running time of SGUnet series and other lightweight networks to process a single MMCBNU_6000 data set image on the NVIDIA embedded platform

Download Citation

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享