Light Field Image Super-Resolution Based on Feature Interaction Fusion and Attention Mechanism

Xinyi Xu; Huiping Deng; Sen Xiang; Jin Wu

doi:10.3788/LOP221911

Laser & Optoelectronics Progress, Volume. 60, Issue 14, 1410017(2023)

Light Field Image Super-Resolution Based on Feature Interaction Fusion and Attention Mechanism

Xinyi Xu^1,2、*, Huiping Deng^1,2, Sen Xiang^1,2, and Jin Wu^1,2

Author Affiliations

¹School of Information Science and Engineering, Wuhan University of Science and Technology, Wuhan 430081, Hubei, China

²Engineering Research Center for Metallurgical Automation and Measurement Technology of Ministry of Education, Wuhan University of Science and Technology, Wuhan 430081, Hubei, China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(13)

Fig. 1. Overall architecture of proposed network

Download full size

Fig. 2. Feature extraction module

Download full size

Fig. 3. Structure diagram of FIAF_module

Download full size

Fig. 4. Feature channel attention module

Download full size

Fig. 5. Structure diagram of SC_module

Download full size

Fig. 6. Comparison of visual effects of different algorithms with ×2 super-resolution

Download full size

Fig. 7. Comparison of visual effects of different algorithms with ×4 super-resolution

Download full size

Fig. 8. Comparison of the visual effects of different algorithms in real-world scenes

Download full size

Table 1. Number of scenarios for training and test sets on 5 datasets
View table
Table 1. Number of scenarios for training and test sets on 5 datasets
Dataset EPFL HCInew HCIold INRIA STFgantry Total
Training 70 20 10 35 9 144
Test 10 4 2 5 2 23

Table 2. Comparison of ×2 super-resolution PSNR/SSIM indicators

View table

Table 2. Comparison of ×2 super-resolution PSNR/SSIM indicators

Algorithm	Scale	EPFL	HCInew	HCIold	INRIA	STFgantry
Bicubic	×2	29.740/0.9376	31.887/0.9356	37.686/0.9785	31.331/0.9577	31.063/0.9498
VDSR	×2	32.498/0.9598	34.371/0.9561	40.606/0.9867	34.439/0.9741	35.541/0.9789
EDSR	×2	33.089/0.9629	34.828/0.9592	41.014/0.9874	34.985/0.9764	36.296/0.9818
RCAN	×2	33.159/0.9634	35.022/0.9603	41.125/0.9875	35.046/0.9769	36.670/0.9831
GB	×2	31.485/0.9564	33.736/0.9523	39.885/0.9853	33.297/0.9713	33.983/0.9719
LFBM5D	×2	31.243/0.9563	33.735/0.9527	40.029/0.9856	33.118/0.9713	33.932/0.9716
LFNet	×2	33.106/0.9684	36.051/0.9707	42.621/0.9920	34.889/0.9789	37.105/0.9873
resLF	×2	33.617/0.9706	36.685/0.9739	43.422/0.9932	35.395/0.9804	38.354/0.9904
LFSSR	×2	33.671/0.9744	36.802/0.9749	43.811/0.9938	35.279/0.9832	37.944/0.9898
LF-ATO	×2	34.166/0.9742	37.080/0.9742	43.640/0.9936	35.985/0.9833	39.582/0.9920
LF-InterNet	×2	34.314/0.9752	36.997/0.9757	44.363/0.9942	35.986/0.9839	38.371/0.9908
MEG-Net	×2	34.332/0.9761	37.186/0.9766	43.965/0.9940	36.049/0.9839	38.678/0.9912
Proposed algorithm	×2	34.493/0.9767	37.468/0.9790	44.252/0.9943	36.366/0.9843	38.894/0.9916

Table 3. Comparison of ×4 super-resolution PSNR/SSIM indicators

View table

Table 3. Comparison of ×4 super-resolution PSNR/SSIM indicators

Algorithm	Scale	EPFL	HCInew	HCIold	INRIA	STFgantry
Bicubic	×4	25.264/0.8324	27.715/0.8517	32.576/0.9344	26.952/0.8867	26.087/0.8452
VDSR	×4	27.246/0.8777	29.308/0.8823	34.810/0.9515	29.186/0.9204	28.506/0.9009
EDSR	×4	27.833/0.8854	29.591/0.8869	35.176/0.9536	29.656/0.9257	28.703/0.9072
RCAN	×4	27.907/0.8863	29.694/0.8886	35.359/0.9548	29.805/0.9276	29.021/0.9131
GB	×4	26.900/0.8737	29.128/0.8817	34.473/0.9523	28.774/0.9183	28.162/0.8953
LFBM5D	×4	26.810/0.8774	29.351/0.8885	34.797/0.9573	28.805/0.9246	28.184/0.9035
LFNet	×4	27.899/0.8991	30.382/0.9067	36.159/0.9657	29.967/0.9381	29.718/0.9322
resLF	×4	28.260/0.9035	30.723/0.9107	36.705/0.9682	30.338/0.9412	30.191/0.9372
LFSSR	×4	28.596/0.9118	30.928/0.9145	36.907/0.9696	30.585/0.9467	30.570/0.9426
LF-ATO	×4	28.514/0.9115	30.880/0.9135	36.999/0.9699	30.711/0.9484	30.607/0.9430
LF-InterNet	×4	28.708/0.9139	30.977/0.9163	37.078/0.9714	30.670/0.9485	30.491/0.9423
MEG-Net	×4	28.656/0.9137	31.002/0.9162	37.109/0.9709	30.590/0.9478	30.619/0.9435
Proposed algorithm	×4	28.701/0.9148	31.131/0.9208	37.210/0.9724	30.734/0.9503	30.610/0.9432

Table 4. Comparison of ×2 super-resolution model parameters and running time (s)

View table

Table 4. Comparison of ×2 super-resolution model parameters and running time (s)

Algorithm	Params /10⁶	EPFL	HCInew	HCIold	INRIA	STFgantry
VDSR	0.665	104.55	40.00	45.47	49.78	20.63
EDSR	38.890	727.34	267.79	299.01	360.05	133.24
RCAN	15.360	178.69	60.14	69.80	82.69	32.31
LFNet	10.385	203.57	84.35	95.92	107.48	48.02
resLF	8.648	749.12	272.48	310.60	368.29	136.84
LFSSR	1.774	79.51	30.30	33.73	37.46	15.00
LF-ATO	1.364	561.25	205.72	233.35	280.79	104.66
LF-InterNet	5.483	193.83	72.06	81.27	96.98	36.79
MEG-Net	1.775	285.45	104.00	119.35	141.54	53.08
Proposed algorithm	4.273	162.43	43.56	48.51	57.70	18.74

Table 5. Comparison of results of ablation experiments
View table
Table 5. Comparison of results of ablation experiments
Model PSNR SSIM
Bicubic 29.740 0.9376
Spatial only 32.853 0.9464
Angular only 30.674 0.9398
Non-FIAF_module 33.887 0.9587
Non-FCA_module 34.185 0.9638
Non-SC_module 34.401 0.9695
All module 34.493 0.9767

Tools

Get Citation

Copy Citation Text

Xinyi Xu, Huiping Deng, Sen Xiang, Jin Wu. Light Field Image Super-Resolution Based on Feature Interaction Fusion and Attention Mechanism[J]. Laser & Optoelectronics Progress, 2023, 60(14): 1410017

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites