(3.235.25.169) 您好!臺灣時間:2021/04/18 04:30
字體大小: 字級放大   字級縮小   預設字形  
回查詢結果

詳目顯示:::

我願授權國圖
: 
twitterline
研究生:王嘉慶
研究生(外文):Chia-Ching Wang
論文名稱:基於深層對抗網絡和感知損失模擬人臉老化
論文名稱(外文):Face aging generated by deep adversarial network and perceptual loss
指導教授:劉宗榮劉宗榮引用關係
指導教授(外文):Tsung-Jung Liu
口試委員:廖俊睿劉冠顯
口試委員(外文):Jan-Ray LiaoKuan-hsien Liu
口試日期:2018-12-19
學位類別:碩士
校院名稱:國立中興大學
系所名稱:電機工程學系所
學門:工程學門
學類:電資工程學類
論文種類:學術論文
論文出版年:2019
畢業學年度:107
語文別:中文
論文頁數:83
中文關鍵詞:年齡老化深度學習機器學習非監督式學習
外文關鍵詞:Face agingdeep learningmachine learningunsupervised learningGAN(Generative Adversarial Network)
相關次數:
  • 被引用被引用:0
  • 點閱點閱:80
  • 評分評分:系統版面圖檔系統版面圖檔系統版面圖檔系統版面圖檔系統版面圖檔
  • 下載下載:0
  • 收藏至我的研究室書目清單書目收藏:0
在論文中,提出了基於深度學習網路架構下的對抗式學習並透過此架構實現人類年齡老化的圖像預測。在事前資料收集,我們會準備年輕人與老年人(目標年齡)的圖片資料集,年輕人的圖片資料集代表輸入圖像(Domain X),老年人的圖片資料集代表輸出(目標年齡)圖像(Domain Y),透過更換不同的老年人(目標年齡)的圖片資料集,我們的網路可以輸出不同的年齡的年輕人的圖像。在對抗式學習的架構中,將會有2個種類的學習網路,分別是用於生成假圖像的生成器與鑑別圖像真實度的鑑別器,在訓練過程中,我們的網路必須遵從賽局理論,也就是說,生成器將會得到鑑別器的反饋資料進而提升自己生成圖像的品質,而鑑別器必須不斷的鑑別假圖像與真圖像之間的差異,當分不出差異時,鑑別器就會以更嚴格的標準來分辨真假圖像,在這中間的訓練,生成器變得優秀,鑑別器也會變得優秀,最終生成器與鑑別器會達到奈許均衡點,而奈許均衡點的目標就是,生成器能夠生成與目標年齡一樣的圖像,而鑑別器辨別生成圖像與目標年齡圖像將會分不出差異,在機率上來說,生成器生成與目標年齡圖片相似的機率會趨近1,而鑑別器鑑別生成圖片與目標年齡圖片的機率會是一樣的,就是辨別出生成圖片與目標年齡的機率都是0.5。在生成器的部分,我們採用深層的ResNet block,根據實驗結果,從4層到7層,經過不同的層數,所產生的圖片品質也有不同,老年化程度的結果表明,7層的結果最優秀。再來是鑑別器的部分,由於產生的圖片過度與(目標年齡)圖片相似,所以會產生一定程度的偽影,為了消除偽影,我們採用了一種名為bias loss的損失函數來降低鑑別器的嚴格程度,而結果也有明顯的改善。在強化老化過程中,我們引入了VGG16的預訓練網路來替我們的Domain X與Domain Y的圖片集取出額外的層數特徵來做損失函數,而這個結果變化是可見的。在圖片資料集,我們取自CACD2000,用於跨年齡的人臉識別和檢索。該數據集包含超過160,000張2000名名人的圖像,年齡介於16至62歲之間。據我們所知,它是迄今為止最大的公開可用的跨年齡人臉數據集。實驗結果表明,該方法可以在我們的數據集以及其他廣泛使用的跨年齡人臉識別數據集MORPH數據集上實現最先進的性能。而我們將不同年齡的人做分類,分別是10~20,20~30,30~40,40~50,50~60,60以上,這6個年齡層,我們使用10~20歲的圖片作為我們的輸入,而後根據想轉換的年齡區間帶入目標年齡的圖片。
In the paper, the adversarial learning based on the deep learning network architecture is proposed and the image prediction of human ageing is realized through this architecture. In the pre-existing data collection, we will prepare a photo data set for young people and the elderly (target age). The young people''s photo data set represents the input image (Domain X), and the elderly photo data set represents the output (target age) map. Like (Domain Y), our network can output images of young people of different ages by changing the image data sets of different seniors (target ages). In the framework of confrontational learning, there will be two kinds of learning networks, namely generators for generating fake images and discriminators for authenticating image realities. During the training process, our network must Compliance with the game theory, that is, the generator will get the feedback data of the discriminator to improve the quality of the generated image, and the discriminator must constantly identify the difference between the fake image and the true image. When the difference occurs, the discriminator will distinguish the true and false images with stricter standards. In the middle of the training, the generator becomes excellent, the discriminator will become excellent, and the generator and discriminator will reach the end. Equilibrium point, and the goal of the Nash Equilibrium point is that the generator can generate the same image as the target age, and the discriminator discriminates between the generated image and the target age image, and in terms of probability, the generator The probability of generating a picture similar to the target age will be close to 1, and the discriminator will have the same chance of identifying the picture and the target age picture, that is, the probability of identifying the generated picture and the target age is 0.5. In the generator part, we use the deep ResNet block. According to the experimental results, from 4 layers to 7 layers, the quality of the pictures produced by different layers is different. The results of the ageing show that the results of the 7 layers are the most excellent. Then there is the part of the discriminator. Since the generated picture is too similar to the (target age) picture, a certain degree of artifact is generated. In order to eliminate the artifact, we use a loss function called bias loss to reduce the discriminator. The degree of rigor, and the results have also improved significantly. In the aging process, we introduced the VGG16 pre-training network to take extra layer features for our Domain X and Domain Y image sets to make a loss function, and the resulting change is visible. In the photo collection, we took CACD2000 for face recognition and retrieval across ages. The dataset contains images of more than 160,000 2,000 celebrities between the ages of 16 and 62. To the best of our knowledge, it is by far the largest publicly available cross-age face dataset. Experimental results show that this method can achieve the most advanced performance on our dataset and other widely used cross-age face recognition dataset MORPH datasets. And we classify people of different ages, which are 10~20, 20~30, 30~40, 40~50, 50~60, 60 or above. For these 6 age groups, we use pictures from 10~20 years old. Our input is then taken into the image of the target age based on the age range we want to convert.
摘要 i
Abstract iii
目錄 v
圖目錄 vi
表目錄 viii
第一章 緒論 1
1.1 前言 1
1.2 論文的動機與大綱 2
第二章 文獻探討與相關背景知識 7
2.1 深度學習 7
2.2 生成對抗網路 12
2.3 年齡老化 15
第三章 論文研究方法 16
3.1 非配對式年齡老化 16
第四章 實驗結果 58
4.1 資料庫介紹 58
4.2 客觀測試分析 65
4.3 主觀測試分析 74
4.4 其他方面的應用 77
第五章 結論與未來展望 80
參考文獻 81
[1] A. P. a. L. W. Y. Taigman, “Unsupervised cross domain image generation,” arXiv preprint arXiv:1611.02200, 2016.
[2] C.-S. C. a. W. H. H. B.-C. Chen, “Face recognition and retrieval using cross-age reference coding with cross-age celebrity dataset,” IEEE TMM, 17(6):804–815, 2015.
[3] G. e. a. Bingham, “Preliminary Studies on a Large Face Database MORPH-II”.
[4] A. D. a. T. Brox, “Generating images with perceptual ceptual similarity metrics based on deep networks,” In NIPS, 2016.
[5] B.-C. C.-S. C. a. W. H. H. Chen, “Face recognition and retrieval using cross-age reference coding with cross-age celebrity dataset,” IEEE Transactions on Multimedia, 2015.
[6] S. S. a. S. M. I. Kemelmacher-Shlizerman, “Illumination-aware age progression,” In CVPR, pages 3334–3341, 2014.
[7] S. C. Z. S. S. a. X. C. J. Suo, “A compositional and dynamic model for face aging,” IEEE PAMI, 2010.
[8] X. C. S. S. W. G. a. Q. D. J. Suo, “ A concatenational graph evolution aging model,” IEEE PAMI, 2012.
[9] Z. A. X. Y. L. L. B. S. a. H. L. S. Reed, “Generative adversarial text to image synthesis,” ICML, 2016.
[10] L. S. M. R. E. S. a. J. B. P. J. T. Todd, “The perception of human growth,” 1980.
[11] J. Z. T. Z. a. A. E. P. Isola, “ Image-to-image translation lation with conditional adversarial networks,” 2016.
[12] C. C. a. Y. L. M. Mathieu, “Deep multi-scalevideo prediction beyond mean square error,” 2015.
[13] G. Qi, “Loss-Sensitive Generative Adversarial Networks on Lipschitz Densities,” arXiv preprint arXiv:1701.06264, 2017.
[14] Y. S. W. W. R. B. D. Z. a. S. Y. S. Liu, “ Face aging with contextural generative adversarial nets.,” In ACM Multimedia, 2017.
[15] Y. S. a. H. Q. Z. Zhang, “Age progression/regression by conditional adversarial autoencoder,” In CVPR, 2017.
[16] Z. C. Y. Y. J. F. S. Y. X. S. a. N. S. W. Wang, “Recurrent face aging,” In CVPR, pages 2378–2386, 2016.
[17] Y. T. a. A. K. J. U. Park, “Age-invariant face recognition nition,” IEEE PAMI, 2010.
[18] Z. Z. W. L. a. F. J. Y. Wang, “ Combining tensor space analysis and active appearance models for aging effect simulation on face images,” IEEE TSMC-B, 42(4):1107–1118, 2012.
[19] J. T. H. L. L. L. a. S. Y. X. Shu, “Personalized age progression with aging dictionary,” In ICCV, p. 3970–3978, 2015.
[20] K. S. a. A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” arXiv:1409.1556, 2014.
[21] ajithvallabai, “Trying to implement code for Learning Face Age Progression : A Pyramid Architecture of GANs,” [線上]. Available: https://github.com/ajithvallabai/Pyramid-Architecture-of-GANs. [存取日期: 12 2018].
[22] N. R. a. R. Chellappa, “Modeling shape and textural variations in aging faces,” In FG, p. 1–8, 2008.
[23] Y. S. G. S. a. X. L. J. Wang, “ Age simulation for face recognition,” In ICPR, volume 3, pages 913–916, 2006.
[24] Q. L. H. X. R. Y. L. Z. W. a. S. P. S. X. Mao, “Least squares generative adversarial networks,” 2016.
[25] J.-Y. e. a. Zhu, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” 2017.
[26] P. Murugan, "Feed forward and backward run in deep convolution neural network," arXiv preprint arXiv:1711.03278, 2017.
[27] I. e. a. Goodfellow, “Generative adversarial nets,” Advances in neural information processing systems, 2014.
[28] X. Z. S. R. a. J. S. K. He, “Deep residual learning for image recognition,” In CVPR, pages 770–778, 2016.
[29] H. e. a. Yang, “Learning face age progression: A pyramid architecture of gans,” 2017.
[30] G. M. B. a. J.-L. D. Antipov, “Face aging with conditional generative adversarial networks,” Image Processing (ICIP), 2017.
[31] J. A. A. a. L. F.-F. Johnson, “Perceptual losses for real-time style transfer and super-resolution,” European Conference on Computer Vision, 2016.
[32] “AgingBooth,” PiVi & Co, [線上]. Available: https://itunes.apple.com/us/app/agingbooth/id357467791?mt=8.
[33] “Face of the future,” Computer Science Dept , [線上]. Available: http://cherry.dcs.aber.ac.uk/Transformer/index.html.
[34] K. R. a. T. Tesafaye, “Morph: A longitudinal image database of normal adult age-progression,” In FG, 2015.
[35] “Face++ Research Toolkit,” Megvii Inc, [線上]. Available: http://www.faceplusplus.com/.
[36] S. C. a. L. B. M. Arjovsky, “Wasserstein GAN,” arXiv preprint arXiv:1701.07875, 2017.
[37] A. Lanitis., “Evaluating the performance of face-aging algorithms,” In FG, pages 1–6,, 2008.
[38] K. G. Q. K. L. N. L. a. M. S. C. Nhan Duong, “Temporal non-volume preserving approach to facialage-progression and age-invariant face recognition,” In ICCV, 2017.
[39] K. L. K. G. Q. a. T. D. B. C. Nhan Duong, “Longitudinal face modeling via temporal deep restricted boltzmann machines,” In CVPR, p. 5772–5780, 2016.
[40] A. V. a. A. Z. O. M. Parkhi, “Deep face recognition,” In BMVC, 2015.
QRCODE
 
 
 
 
 
                                                                                                                                                                                                                                                                                                                                                                                                               
第一頁 上一頁 下一頁 最後一頁 top
系統版面圖檔 系統版面圖檔