Cycle generative adversarial network guided by dual special attention mechanism

Fig. 6. Effect of the generated image’‍s quality that with or without our generator，discriminator and cycle consistency loss function background mask.（a）Original image；（b）Generator without special attention-mechanism guided；（c）Discriminator without special attention-mechanism guided；（d）Without cycle consistency loss function of background mask；（e）All of three factors.

Download full size

View in Article

Fig. 7. Effect of different λ-factors on the quality of the generated images on the horse-to-zebra task

Download full size

View in Article

Fig. 8. Attention mask and images generated by ours model on the apple to orange and orange to apple tasks

Download full size

View in Article

Fig. 9. Attention mask and images generated by ours model on the horse to zebra and zebra to horse tasks

Download full size

View in Article

Fig. 10. Performance of different models on the horse-zebra interchange task.（a）Original image；（b）CycleGAN；（c）RA；（d）DiscoGAN；（e）UNIT；（f）DualGAN；（g）UAIT；（h）AttentionGAN；（i）Ours.

Download full size

View in Article

Fig. 11. Performance of different models on the selfie-anime interchange task.（a）Original image；（b）CycleGAN；（c）UNIT；（d）MUNIT；（e）DRIT；（f）U-GAT-IT；（g）AttentionGAN；（h）Ours.

Download full size

View in Article

Table 1. Details of each dataset
View table
View in Article
Table 1. Details of each dataset
数据集训练集/张测试集/张总计/张
Horse2Zerba^［12］ 2 041 260 2 661
Apple2Orange^［12］ 2 016 514 2 530
Selfie2Anime^［17］ 6 800 200 7 000

Table 2. Ablation studies of models on the horse-to-zebra task
View table
View in Article
Table 2. Ablation studies of models on the horse-to-zebra task
设置无注意力引导生成器无注意力引导鉴别器无背景掩码循环一致性损失三者都有
FID 107.23 92.24 86.05 57.54

Table 3. Effect of different λ‍-coefficients on the quality of the generated images
View table
View in Article
Table 3. Effect of different λ‍-coefficients on the quality of the generated images
设置 FID
$λ = 0.0$ 86.05
$λ = 1.0$ （本文） 57.54
$λ = 10.0$ 79.58

Table 4. $K I D \times 100 \pm s t d . \times 100$ metrics for different models on different tasks

View table

View in Article

Table 4. $K I D \times 100 \pm s t d . \times 100$ metrics for different models on different tasks

模型	苹果转橘子	橘子转苹果	马转斑马	斑马转马
DiscoGAN^［8］	$18.34 \pm 0.75$	$21.56 \pm 0.80$	$13.68 \pm 0.75$	$16.60 \pm 0.50$
RA^［19］	$12.75 \pm 0.49$	$13.84 \pm 0.78$	$10.16 \pm 0.12$	$10.97 \pm 0.26$
DualGAN^［7］	$13.04 \pm 0.72$	$12.42 \pm 0.88$	$10.38 \pm 0.31$	$12.86 \pm 0.50$
UNIT^［9］	$11.68 \pm 0.43$	$11.76 \pm 0.51$	$11.22 \pm 0.24$	$13.63 \pm 0.34$
CycleGAN^［12］	$8.48 \pm 0.53$	$9.82 \pm 0.51$	$10.25 \pm 0.25$	$11.44 \pm 0.38$
UAIT^［13］	$6.44 \pm 0.69$	$5.32 \pm 0.48$	$6.93 \pm 0.27$	$8.87 \pm 0.26$
AttentionGAN^［16］	$10.03 \pm 0.66$	$4.38 \pm 0.42$	$2.03 \pm 0.64$	$6.48 \pm 0.51$
Ours	$7.28 \pm 0.71$	$3.02 \pm 0.29$	$1.13 \pm 0.23$	$5.06 \pm 0.60$

Table 5. ‍KID $\times$ 100 $\pm$ std. $\times$ 100 metrics for different models on selfie to anime task
View table
View in Article
Table 5. ‍KID $\times$ 100 $\pm$ std. $\times$ 100 metrics for different models on selfie to anime task
模型自拍转动漫
CycleGAN^［12］ $13.08 \pm 0.49$
UNIT^［9］ $14.71 \pm 0.59$
MUNIT^［10］ $13.85 \pm 0.41$
DRIT^［11］ $15.08 \pm 0.62$
U-GAT-IT^［17］ $11.61 \pm 0.57$
AttentionGAN^［16］ $12.14 \pm 0.43$
Ours $10.01 \pm 0.29$

Table 6. FID metrics for different models on horse to zebra task
View table
View in Article
Table 6. FID metrics for different models on horse to zebra task
模型马转斑马
UNIT^［9］ $241.13$
CycleGAN^［12］ $109.36$
DA-GAN^［20］ $103.42$
TransGaGa^［21］ $95.81$
SAT^［15］ $98.90$
AttentionGAN^［16］ $68.55$
Ours $57.54$

Table 7. Complexity analysis of different models guided by attention mechanism

View table

View in Article

Table 7. Complexity analysis of different models guided by attention mechanism

模型	参数量/M	浮点运算量/G	乘法累加运算量/G	训练时间（100张）/s	占用显存（1批量）/MiB
X.‍ Chen^［14］	43.94	102.01	204.02	55	5 069
T. Hao^［16］	29.22	74.67 G	149.31	47	5 463
Ours	29.49	65.50	131.00	36	3 669

Tools

Get Citation

Copy Citation Text

Jun-ming LAO, Wu-jian YE, Yi-jun LIU, Kai-yi YUAN. Cycle generative adversarial network guided by dual special attention mechanism[J]. Chinese Journal of Liquid Crystals and Displays, 2022, 37(6): 746

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category:

Received: Nov. 24, 2021

Accepted: --

Published Online: Jun. 22, 2022

The Author Email: Wu-jian YE (yewjian@126.com)

DOI:10.37188/CJLCD.2021-0293

Topics

Table 1. Details of each dataset

Table 1. Details of each dataset

Table 2. Ablation studies of models on the horse-to-zebra task

Table 2. Ablation studies of models on the horse-to-zebra task

Table 3. Effect of different λ‍-coefficients on the quality of the generated images

Table 3. Effect of different λ‍-coefficients on the quality of the generated images

Table 4. KID×100±std. ×100 metrics for different models on different tasks

Table 4. KID×100±std. ×100 metrics for different models on different tasks

Table 5. ‍KID×100±std.×100 metrics for different models on selfie to anime task

Table 5. ‍KID×100±std.×100 metrics for different models on selfie to anime task

Table 6. FID metrics for different models on horse to zebra task

Table 6. FID metrics for different models on horse to zebra task

Table 7. Complexity analysis of different models guided by attention mechanism

Table 7. Complexity analysis of different models guided by attention mechanism

Table 4. $K I D \times 100 \pm s t d . \times 100$ metrics for different models on different tasks

Table 4. $K I D \times 100 \pm s t d . \times 100$ metrics for different models on different tasks

Table 5. ‍KID $\times$ 100 $\pm$ std. $\times$ 100 metrics for different models on selfie to anime task

Table 5. ‍KID $\times$ 100 $\pm$ std. $\times$ 100 metrics for different models on selfie to anime task