Acta Optica Sinica, Volume. 42, Issue 5, 0533002(2022)
Progressive Multi-Scale Feature Cascade Fusion Color Constancy Algorithm
Figures & Tables(13)
Fig. 2. Features extracted from different network structures. (a) Input image; (b) features extracted by algorithm in this pape; (c) features extracted by series convolution network
Fig. 5. Comparison of color constancy algorithms under complex scene illumination. (a) Input images; (b) grey-world algorithm; (c) grey-edge algorithm; (d) FC4 algorithm; (e) C4 algorithm; (f) proposed algorithm
Fig. 9. Visualization results of network tests. (a) Input images; (b) images after light source correction obtained by proposed algorithm; (c) features extracted by proposed algorithm; (d) C4 algorithm;(e) FC4 algorithm;(f) grey-world algorithm; (g) white-patch algorithm; (h) grey-edge algorithm; (i) SOG algorithm
Table 1. Structures of convolution branch networks
View tableTable 1. Structures of convolution branch networks
Branch Conv1 Conv2 Conv3 Conv4 Conv5 1 3×3×64 -- -- -- -- 2 5×5×128 1×1×64 5×5×64 -- -- 3 7×7×64 5×5×128 1×1×64 5×5×128 3×3×64
Table 2. Layer numbers of convolution branch network
View tableTable 2. Layer numbers of convolution branch network
Branch Conv1 Conv2 Conv3 Conv4 Conv5 Conv6 Conv7 1 3×3×64 -- -- -- -- -- -- 2 5×5×128 1×1×64 5×5×64 -- -- -- -- 3 7×7×64 5×5×128 1×1×64 5×5×128 3×3×64 -- -- 4 7×7×64 5×5×128 3×3×64 3×3×128 1×1×64 3×3×128 1×1×64
Table 3. Test results obtained by using reprocessed ColorChecker dataset
View tableTable 3. Test results obtained by using reprocessed ColorChecker dataset
Method Mean Med Tri Best25% Worst25% 95th Grey-world[ 21 ]10.700 10.600 10.700 3.450 12.300 17.400 White-patch[ 25 ]9.800 8.000 8.900 3.800 13.600 22.300 Shades-of-grey[ 26 ]8.300 7.500 7.800 2.900 11.800 17.000 1-order grey edge[ 4 ]5.000 3.700 4.100 3.900 10.100 13.300 2-order grey edge[ 4 ]5.400 4.500 4.800 2.600 9.800 12.800 Pixel-based gamut[ 27 ]6.900 5.200 5.700 1.800 11.700 18.200 Edge-based gamut[ 27 ]6.900 4.600 5.200 2.100 14.600 20.600 Bayesian[ 6 ]6.600 4.600 5.200 3.200 10.900 18.400 General grey-world[ 28 ]7.600 6.700 7.000 3.800 12.100 16.500 Using CNNs[ 7 ]8.200 6.300 6.800 2.600 11.300 20.400 Deep color constancy[ 19 ]5.700 4.700 5.000 3.200 8.400 12.400 Exemplar[ 27 ]2.890 2.270 2.420 0.820 5.970 6.950 CCC[ 29 ]2.000 1.220 1.400 0.350 4.760 5.850 CC-GANs (pix-pix)[ 13 ]3.600 2.800 3.100 1.200 7.200 9.400 FC4-AlexNet[ 14 ]1.770 1.110 1.290 0.340 4.290 5.440 FC4-SqueezeNet[ 14 ]1.650 1.180 1.270 0.380 3.780 4.730 IEN+PSN[ 18 ]2.250 1.590 1.730 0.590 5.030 6.080 Multi-hypothesis[ 30 ]2.100 1.320 1.530 0.360 5.100 -- C4[ 22 ]1.350 0.880 0.990 0.280 3.210 -- Our method 1.210 0.767 0.843 0.244 2.655 3.043
Table 4. Test results obtained by using NUS-8 dataset
View tableTable 4. Test results obtained by using NUS-8 dataset
Method Mean Med Tri Best25% Worst25% Grey-word[ 21 ]4.140 3.200 3.390 0.900 9.000 White-patch[ 25 ]10.620 10.580 10.490 1.860 19.450 Shades-of-grey[ 26 ]3.400 2.570 2.730 0.770 7.410 1-order grey edge[ 4 ]3.200 2.220 2.430 0.720 7.360 2-order grey edge[ 4 ]3.200 2.260 2.440 0.750 7.270 Pixel-based gamut[ 27 ]7.700 6.710 6.900 2.510 14.050 Edge-based gamut[ 27 ]8.430 7.050 7.370 2.410 16.080 Bayesian[ 6 ]3.670 2.730 2.910 0.820 8.210 Using CNNs[ 7 ]7.600 6.900 7.400 3.000 12.400 Deep color constancy[ 18 ]6.200 5.000 5.400 3.900 8.600 CCC[ 29 ]2.800 1.800 1.900 0.850 6.300 CC-GANs (pix-pix)[ 13 ]3.800 3.000 3.700 1.900 8.400 FC4-AlexNet[ 14 ]2.120 1.530 1.670 0.480 4.780 FC4-SqueezeNet[ 14 ]2.230 1.570 1.270 0.470 5.150 IEN+PSN[ 18 ]2.100 1.350 1.510 0.450 5.010 Multi-hypothesis[ 30 ]2.350 1.550 1.730 0.460 5.620 C4[ 22 ]1.960 1.420 1.530 0.480 4.400 Our method 1.574 1.027 1.241 0.386 3.512
Tools
Get Citation
Copy Citation Text
Zepeng Yang, Kai Xie, Tong Li. Progressive Multi-Scale Feature Cascade Fusion Color Constancy Algorithm[J]. Acta Optica Sinica, 2022, 42(5): 0533002
Paper Information
Category: Vision, Color, and Visual Optics
Received: Aug. 17, 2021
Accepted: Sep. 23, 2021
Published Online: Apr. 17, 2022
The Author Email: Xie Kai (2596898130@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20