Multi-exposure Image Fusion Based on Attention Mechanism

Bendu BAI and Junpeng LI^*

Author Affiliations

School of Communication and Information Engineering，Xi'an University of Posts & Telecommunications，Xi'an 710121，China

show less

DOI: 10.3788/gzxb20225104.0410004 Cite this Article

Bendu BAI, Junpeng LI. Multi-exposure Image Fusion Based on Attention Mechanism[J]. Acta Photonica Sinica, 2022, 51(4): 0410004 Copy Citation Text

show less

Fig. 1. Network structure of AMEFNet

Fig. 2. Feature extraction network structure of AMEFNet

Fig. 3. Attention network structure of AMEFNet

Fig. 4. Feature reconstruction network structure of Unet

Fig. 5. Loss function curve of AMEFNet

Fig. 6. Comparison of House sequence algorithm results

Fig. 7. Comparison of TableLamp sequence algorithm results

Fig. 8. Comparison of running time

Fig. 9. Exposure fusion results of Studio sequence

Fig. 10. Comparison of running time

Fusion Method	PSNR	AG	SF	EN	VIF
GFF	58.224 0	5.623 4	18.826 9	7.393 9	0.825 3
DSIFT	59.825 6	5.105 1	17.103 8	7.354 3	0.759 1
SPD-MEF	58.754 5	5.877 3	21.033 9	7.039 9	0.794 9
MEFNet	58.380 4	6.029 3	20.596 6	7.394 8	0.844 9
MEF-GAN	58.773 6	4.756 5	13.554 8	6.982 3	0.621 1
Ours	59.115 0	6.199 1	21.098 6	7.203 7	0.874 6

Table 1. The values of five quality metrics averaged over the fused images on test set

Table 2. Time complexity of the different algorithm

Fusion method	PSNR	AG	SF	EN	VIF
RAMEFNet	58.750 9	5.986 6	20.135 1	6.9819	0.833 9
AMEFNet	59.115 0	6.199 1	21.098 6	7.203 7	0.874 6

Table 3. Evaluations of attention module on five quality metrics average

Fusion method	PSNR	AG	SF	EN	VIF
AMEFNet- $λ_{0.02}$	59.249 8	6.035 1	20.113 4	7.1011	0.843 1
AMEFNet- $λ_{0.2}$	59.115 0	6.199 1	21.098 6	7.203 7	0.874 6
AMEFNet- $λ_{2}$	59.0990	6.174 3	21.035 9	7.1915	0.879 5