Image Super-Resolution Reconstruction Based on Multi-Scale Residual Aggregation Feature Network

Lifeng He; Liangliang Su; Guangbin Zhou; Pu Yuan; Bofan Lu; Jiajia Yu

doi:10.3788/LOP202158.2410011

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 24 >Page 2410011 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 24, 2410011 (2021)

Image Super-Resolution Reconstruction Based on Multi-Scale Residual Aggregation Feature Network

Lifeng He^1、2, Liangliang Su^1、*, Guangbin Zhou¹, Pu Yuan¹, Bofan Lu¹, and Jiajia Yu¹

Author Affiliations

¹School of Electronic Information and Artificial Intelligence, Shaanxi University of Science & Technology, Xi'an, Shaanxi 710021, China;

²School of Information Science and Technology, Aichi Prefectural University, Nagakute, Aichi 480- 1198, Japan

show less

DOI: 10.3788/LOP202158.2410011 Cite this Article Set citation alerts

Lifeng He, Liangliang Su, Guangbin Zhou, Pu Yuan, Bofan Lu, Jiajia Yu. Image Super-Resolution Reconstruction Based on Multi-Scale Residual Aggregation Feature Network[J]. Laser & Optoelectronics Progress, 2021, 58(24): 2410011 Copy Citation Text

show less

Fig. 1. Structure of multi-scale residual aggregation feature network

Download full size

Fig. 2. Multi-scale feature extraction module

Download full size

Fig. 3. Extended convolution kernel with different expansion coefficients

Download full size

Fig. 4. Comparison of different module structures. (a) Ordinary residual block; (b) hybrid extended convolution residual block

Download full size

Fig. 5. Gridding artifact with a single pixel convolved with a 3×3 extended convolutional kernel (expansion coefficient r=2)

Download full size

Fig. 6. Diagram of visual feature output. (a) RGB image; (b) without per-pixel addition operation; (c) with per-pixel addition operation

Download full size

Fig. 7. Reconstruction results of the three models in the Urban100 image “img091”. (a) Original drawing; (b) M_HERB; (c) M_RB+AM; (d) M_HERB+AM

Download full size

Fig. 8. Comparison of image reconstruction effects under various methods

Download full size

N₂	2	3	4	5	6	7	8	9	10	11
PSNR /dB	28.335	28.398	28.428	28.444	28.463	28.490	28.522	28.532	28.507	28.506
Average time /s	0.107	0.138	0.176	0.195	0.216	0.235	0.254	0.277	0.299	0.320

Table 1. Relationship between number of hybrid extended convolution residual blocks, average time, and PSNR

N₁	1	2	3	4
PSNR /dB	28.431	28.522	28.527	28.533
Average time /s	0.171	0.254	0.341	0.438

Table 2. Relationship between number of multi-scale feature extraction modules, average time , and PSNR

Data set	M_HERB	M_RB+AM	M_HERB+AM
Set5	32.11/0.8938	32.01/0.8930	32.03/0.8933
Set14	28.49/0.7797	28.41/0.7756	28.52/0.7805
BSD100	27.53/0.7351	27.45/0.7312	27.57/0.7361
Manga109	30.17/0.9055	30.10/0.9011	30.30/0.9072
UrBan100	25.85/0.7792	25.77/0.7789	25.99/0.7846

Table 3. Average PSNR/SSIM of three models on 5 data sets PSNR unit: dB

Method	r'	Set5	Set14	BSD100	Manga109	UrBan100
Bicubic	2	33.66/0.9299	30.24/0.8688	29.56/0.8431	30.80/0.9339	26.88/0.8403
SRCNN^[4]	2	36.66/0.9542	32.45/0.9067	31.36/0.8879	35.60/0.9663	29.50/0.8946
VDSR^[8]	2	37.53/0.9590	33.05/0.9130	31.90/0.8960	37.22/0.9750	30.77/0.9140
DRRN^[10]	2	37.74/0.9591	33.23/0.9136	32.05/0.8973	37.60/0.9736	31.23/0.9188
SRMDNF^[26]	2	37.79/0.9601	33.32/0.9159	32.05/0.8985	38.07/0.9761	31.33/0.9204
IMRSR^[12]	2	37.78/0.9643	33.26/0.8488	32.00/0.9073		31.00/0.9235
Proposed method	2	37.89/0.9603	33.41/0.9159	32.07/0.8986	38.24/0.9763	31.62/0.9237
Bicubic	3	30.39/0.8682	27.55/0.7742	27.21/0.7385	26.95/0.8556	24.46/0.7349
SRCNN^[4]	3	32.75/0.9090	29.30/0.8215	28.41/0.7863	30.48/0.9117	26.24/0.7989
VDSR^[8]	3	33.67/0.9210	29.78/0.8320	28.83/0.7990	32.01/0.9340	27.14/0.8290
DRRN^[10]	3	34.03/0.9244	29.96/0.8349	28.95/0.8004	32.42/0.9359	27.53/0.8378
SRMDNF^[26]	3	34.12/0.9254	30.04/0.8382	28.97/0.8025	33.00/0.9403	27.57/0.8398
IMRSR^[12]	3	33.91/0.9312	29.88/0.8488	28.80/0.8166		27.00/0.8403
Proposed method	3	34.18/0.9255	30.16/0.8389	28.99/0.8033	33.01/0.9413	27.77/0.8450
Method	r'	Set5	Set14	BSD100	Manga109	UrBan100
Bicubic	4	28.42/0.8104	26.00/0.7027	25.96/0.6675	24.89/0.7866	23.14/0.6577
SRCNN^[4]	4	30.48/0.8628	27.50/0.7513	26.90/0.7101	27.58/0.8555	24.52/0.7221
VDSR^[8]	4	31.35/0.8830	28.02/0.7680	27.29/0.7726	28.83/0.8870	25.18/0.7540
DRRN^[10]	4	31.68/0.8888	28.21/0.7721	27.38/0.7284	29.18/0.8914	25.44/0.7638
SRMDNF^[26]	4	31.96/0.8925	28.35/0.7787	27.49/0.7337	30.09/0.9024	25.68/0.7731
IMRSR^[12]	4	31.59/0.8957	28.19/0.7892	27.30/0.7469		25.15/0.7714
Proposed method	4	32.03/0.8933	28.52/0.7805	27.57/0.7361	30.30/0.9072	25.99/0.7846

Table 4. Average PSNR/SSIM of different methods on different test sets PSNR unit: dB

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information