Multi-scale recurrent attention network for image motion deblurring

Xiangjun Wang; Wensen Ouyang

doi:10.3788/IRLA20210605

Journals >Infrared and Laser Engineering >Volume 51 >Issue 6 >Page 20210605 > Article

Infrared and Laser Engineering
Vol. 51, Issue 6, 20210605 (2022)

Multi-scale recurrent attention network for image motion deblurring

Xiangjun Wang^1、2 and Wensen Ouyang^1、2

Author Affiliations

¹State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University, Tianjin 300072, China

²MOEMS Education Ministry Key Laboratory, Tianjin University, Tianjin 300072, China

show less

DOI: 10.3788/IRLA20210605 Cite this Article

Xiangjun Wang, Wensen Ouyang. Multi-scale recurrent attention network for image motion deblurring[J]. Infrared and Laser Engineering, 2022, 51(6): 20210605 Copy Citation Text

show less

Fig. 1. Multi-scale recurrent Neural Network Architecture

Download full size | View in the Article

Fig. 2. Encode，decode block. (a) Encode block; (b) Decode block; (c) End decode block; (d) Residual dense attention block

Download full size | View in the Article

Fig. 3. (a), (b) Convolutional block attention submodule^[15]; (c) Improved CBAM (CBAM-J)

Download full size | View in the Article

Fig. 4. CBAM connection modes in literature[15]

Download full size | View in the Article

Fig. 5. Test result on Lai real blur dataset

Download full size | View in the Article

Fig. 6. Test result on GoPro testing set

Download full size | View in the Article

CBAM connections	Output (y)
Proposed(CBAM-J)	$ x \cdot S(x) \cdot \left( {1+C\left( {x \cdot S(x)} \right)} \right) $
Channel+Spatial	$ x \cdot C(x) \cdot S\left( {x \cdot C(x)} \right) $
Spatial+Channel	$ x \cdot S(x) \cdot C\left( {x \cdot S(x)} \right) $
Spatial // Channel	$ x \cdot S(x)+x \cdot C(x) $

Table 1. Connection modes of CBAM

View in the Article

Method	GoPro		Lai		Köhler
Method	SSIM	PSNR/dB	SSIM	PSNR/dB	SSIM	PSNR/dB
Proposed method	0.9185	29.0284	0.6674	16.5491	0.7625	19.9943
DeblurGAN	0.8474	25.0200	0.6425	15.8905	0.7447	19.7570
DeblurGAN-v2(Inception)	0.9141	28.2701	0.6514	16.1121	0.7469	19.4994
DeblurGAN-v2(MobileNet)	0.8731	25.9644	0.6598	16.4073	0.7556	19.7882
SRN deblur net	0.9331	30.1513	0.6494	16.1000	0.7505	19.5238

Table 2. Deblurring evaluation results on three datasets

View in the Article

Method	FLOPs/G	Size/MB	Time/s
Proposed method	261.19	12.3	0.206
DeblurGAN	678.29	45.6	0.694
DeblurGAN-v2(Inception)	411.34	244.7	0.212
DeblurGAN-v2(MobileNet)	43.75	13.6	0.068
SRN deblur net	1434.82	78.7	0.501

Table 3. Number of parameters & execution time per frame

Xiangjun Wang, Wensen Ouyang. Multi-scale recurrent attention network for image motion deblurring[J]. Infrared and Laser Engineering, 2022, 51(6): 20210605

Download Citation

Tools

Save the article for my favorites

Paper Information