Heterogeneous Noise Iris Segmentation Based on Attention Mechanism and Dense Multiscale Features

Xuanang You; Peng Zhao; Xiaodong Mu; Kun Bai; Sai Lian

doi:10.3788/LOP202259.0410006

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 4 >Page 0410006 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 4, 0410006 (2022)

Heterogeneous Noise Iris Segmentation Based on Attention Mechanism and Dense Multiscale Features

Xuanang You^1、*, Peng Zhao^1、**, Xiaodong Mu¹, Kun Bai¹, and Sai Lian²

Author Affiliations

¹College of Operational Support, Rocket Force University of Engineering, Xi'an , Shaanxi 710025, China

²College of Microelectronics, Xi'an Jiaotong University, Xi'an , Shaanxi 710049, China

show less

DOI: 10.3788/LOP202259.0410006 Cite this Article Set citation alerts

Xuanang You, Peng Zhao, Xiaodong Mu, Kun Bai, Sai Lian. Heterogeneous Noise Iris Segmentation Based on Attention Mechanism and Dense Multiscale Features[J]. Laser & Optoelectronics Progress, 2022, 59(4): 0410006 Copy Citation Text

show less

Examples of eye image acquisition in complex scenes. (a) Gaze deviation; (b) absence of iris; (c) eyelash occlusion; (d) iris rotation; (e) blur; (f) hair shade; (g) specular reflection; (h) glasses occlusion

Fig. 1. Examples of eye image acquisition in complex scenes. (a) Gaze deviation; (b) absence of iris; (c) eyelash occlusion; (d) iris rotation; (e) blur; (f) hair shade; (g) specular reflection; (h) glasses occlusion

Download full size

Fig. 2. MFFIris-Unet architecture

Download full size

Fig. 3. Structure of inverted residual block

Download full size

Fig. 4. Structure of the modified residual bottleneck unit

Download full size

Fig. 5. Spatial-channel parallel attention module architecture

Download full size

Fig. 6. Improved Dense-ASPP structure

Download full size

Fig. 7. Examples of data enhanced training samples

Download full size

Fig. 8. Curves of training loss function and precision change at different datasets. (a) CASIA; (b) UBIRIS; (c) MICHE

Download full size

Fig. 9. Segmentation results of different methods on MICHE dataset. (a) Original image; (b) ground truth; (c) results of Deeplab V3; (d) results of U-Net; (e) results of RTV-L; (f) results of PI-Unet; (g) results of MFFIris-Unet

Download full size

Fig. 10. Segmentation results of different methods on CASIA dataset. (a) Original image; (b) ground truth; (c) results of Deeplab V3; (d) results of U-Net; (e) results of RTV-L; (f) results of PI-Unet; (g) results of MFFIris-Unet

Download full size

Fig. 11. Segmentation results of different methods on UBIRIS dataset. (a) Original image; (b) ground truth; (c) results of Deeplab V3; (d) results of U-Net; (e) results of RTV-L; (f) results of PI-Unet; (g) results of MFFIris-Unet

Download full size

Fig. 12. Histograms of mIoU and average F1 scores on three datasets

Download full size

Fig. 13. Visualized results predicted by the base model and MFFIris-Unet. (a) Original image; (b) ground truth; (c) results of base method; (d) results of MFFIris-Unet

Download full size

Method	Dataset	R		P		F1-Score		mIoU /%	Average time /s
Method	Dataset	μ /%	σ /%	μ /%	σ /%	μ /%	σ /%	mIoU /%	Average time /s
Deeplab V3	CASIA	90.13	6.62	92.80	4.12	93.21	3.67	88.21	0.56
	UBIRIS	85.17	9.53	90.92	4.01	87.55	6.32	79.24	0.44
	MICHE	89.84	10.93	91.66	8.12	91.18	8.89	84.69	0.41
U-Net	CASIA	91.77	7.62	95.23	3.51	91.78	5.58	87.34	0.93
	UBIRIS	91.96	7.82	90.29	4.63	90.81	4.92	81.92	0.67
	MICHE	88.86	13.13	90.75	8.56	88.25	10.52	81.20	0.66
RTV-L	CASIA	80.95	6.59	95.83	3.91	87.55	4.58	78.11	2.68
	UBIRIS	88.23	9.66	85.16	10.58	85.97	8.72	74.01	1.15
	MICHE	84.56	17.61	74.27	16.82	77.10	14.71	64.21	1.57
PI-Unet	CASIA	93.11	8.69	95.22	5.31	96.53	5.41	94.21	0.18
	UBIRIS	91.87	7.43	91.98	4.55	95.25	6.25	92.31	0.26
	MICHE	93.52	10.11	93.65	9.16	94.02	8.52	93.53	0.33
MFFIris-Unet	CASIA	92.62	7.65	96.56	3.69	97.14	4.36	94.61	0.11
	UBIRIS	92.87	6.87	92.96	3.68	96.59	4.11	94.28	0.10
	MICHE	94.05	10.21	93.14	8.69	96.54	8.62	93.63	0.07

Table 1. Evaluation index results of different methods on three iris datasets

Method	Params /10⁶	FLOPs /10⁹	Storage space /GB
FCN8s	134.27	84.99	0.513
U-Net	26.36	62.61	0.121
SegNet	16.31	53.93	0.123
Deeplab V3	18.86	26.29	0.072
PI-Unet	2.86	1.56	0.012
MFFIris-Unet	1.45	0.35	0.005

Table 2. Comparison of the number of parameters, computation amount, and storage space of different methods

Method	Dataset	F1-Score /%	mIoU /%	Average time /s	Train time /h	Model size /MB
Base	CASIA	96.95	94.40	0.76	6	9.69
	UBIRIS	96.23	93.61	0.61	22
	MICHE	96.47	93.51	0.55	23
Base+RBU+Atten	CASIA	96.56	93.85	0.22	3	5.32
	UBIRIS	95.22	93.44	0.16	8
	MICHE	95.62	93.21	0.13	5
Base+RBU+FPM	CASIA	96.89	94.59	0.12	5.5	5.66
	UBIRIS	96.29	93.96	0.08	21
	MICHE	96.54	93.69	0.10	22
MFFIris-Unet	CASIA	97.14	94.61	0.11	3	5.66
	UBIRIS	96.59	94.28	0.07	8
	MICHE	96.54	93.63	0.10	5

Table 3. Results of ablation experiments

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information