Low-light image enhancement algorithm based on multi-channel fusion attention network

Fig. 2. Channel Attention and Spatial Attention Module in CBAM

Fig. 3. Overall network structure of this paper

Fig. 4. Multi-level feature extraction module

Fig. 5. Cross-scale feature attention module

Fig. 6. Subjective visual contrast of different algorithms on synthesizing low-light images

Fig. 7. Subjective visual comparison of different algorithms on real low-light images

Fig. 8. Subjective visual contrast of enhancement effects of different network structures

Fig. 9. Subjective visual contrast of the enhancement effects of different loss functions

Table 1. Average value of evaluation indicators for different algorithms on synthesizing low-light images

Table 1. Average value of evaluation indicators for different algorithms on synthesizing low-light images

方法	HE	SSR	RetinexNet	KinD	Zero-DCE	MBLLEN	RISSNet	本文算法
PSNR↑	13.661 1	9.137 5	18.280 4	21.779 5	22.978 8	25.986 4	27.341 6	28.346 5
SSIM↑	0.660 8	0.564 2	0.816 4	0.825 9	0.849 1	0.874 0	0.884 2	0.906 5
MSE↓	0.037 6	0.084 3	0.023 3	0.006 2	0.004 8	0.002 9	0.001 2	0.000 7
VIF↑	0.237 5	0.410 1	0.519 2	0.536 7	0.571 2	0.578 5	0.631 0	0.682 0
IE↑	7.771 5	6.158 1	6.444 0	6.642 1	6.658 5	6.767 3	0.671 6	6.962 3

Table 2. The average value of evaluation indicators for different algorithms on real low-light images

Table 2. The average value of evaluation indicators for different algorithms on real low-light images

方法	HE	SSR	RetinexNet	KinD	Zero-DCE	MBLLEN	RISSNet	本文算法
PSNR↑	14.056 9	10.359 3	19.853 3	21.468 5	22.250 1	26.046 8	27.001 6	27.978 1
SSIM↑	0.712 8	0.593 2	0.812 8	0.840 9	0.853 7	0.888 5	0.889 2	0.925 5
MSE↓	0.041 2	0.103 3	0.028 9	0.018 3	0.011 3	0.003 5	0.001 3	0.000 7
VIF↑	0.301 3	0.308 2	0.440 2	0.480 6	0.483 3	0.576 0	0.634 6	0.669 9
IE↑	7.944 7	6.826 6	6.607 4	6.799 3	6.710 8	6.856 7	6.876 7	6.984 2

Table 3. The average value of the evaluation index for the enhancement effect of different network structures

Table 3. The average value of the evaluation index for the enhancement effect of different network structures

模型	PSNR↑	SSIM↑	VIF↑
无OctConv	25.091 5	0.890 1	0.499 7
无CBAM	24.633 8	0.892 5	0.505 1
CBAM替换为SE	24.064 1	0.879 1	0.489 2
CBAM替换为ECA	26.318 2	0.893 7	0.536 6
通道数均为16	26.553 4	0.899 8	0.515 4
通道数均为32	25.911 7	0.899 9	0.497 7
通道数均为64	24.918 5	0.891 0	0.509 9
本文算法	27.978 1	0.925 5	0.669 9

Table 4. Average of the evaluation indicators for loss functions with different weight coefficients