臺灣博碩士論文加值系統

隨著近年來人工智慧的發展，機器學習所能應用的領域越來越廣泛，在這其中，尤以深度學習這塊最為突出，並且已成為近年來機器學習領域的主流，無論在圖像生成、生物識別、語意辨識…等，皆有相當優秀的表現，是個能夠廣泛應用於人工智慧各項領域的主流技術。
本篇論文關注於卡通圖像的自動生成，提出了一個應用於組合式圖像生成的部位區域分群系統。在圖像生成領域方面，近年來大多論文所使用的圖像生成模型都是基於深度學習，像是Generative Adversarial Network（GAN）、Variational autoencoder（VAE）、PixelCNN…等，其中GAN更是近兩年來的生成模型主流。這類基於深度學習的圖像生成模型其生成能力皆為相當優秀，但通常需要大量的訓練資料以及較長的運算時間，所需的運算設備也較為昂貴。對於一般大眾使用者來說，通常得仰賴於使用他人所訓練好的單一類別生成模型來進行創作，而無法隨意的依照個人喜好進行多種類別的圖像創作。
本篇論文所提出的部位區域（Region）分群系統是為了應用於組合式卡通圖像生成，先以預訓練的卷積神經網路模型提取輸入圖像部位特徵，再使用淺層網路評估特徵群數並以非監督式學習的方式來對其進行分群，故運算成本以及資料量需求與深度學習相比皆為較低，且不須任何樣本標記資訊。透過降低對訓練資料集的需求，使圖像生成系統能更加容易地達到多類別圖像生成。在實驗結果中顯示，本系統確實能自動評估出較好的分群群數並得到良好的分群結果。

With the development of artificial intelligence in recent years, machine learning can be applied to more and more fields. Among them, deep learning is the most prominent, and has become the mainstream of machine learning in recent years.
This paper focuses on the automatic generation of cartoon images, and proposes a region clustering system for combined image generation. In the area of image generation, the image generation models used in most of the papers in recent years are based on deep learning, such as Generative Adversarial Network (GAN), Variational autoencoder (VAE), etc. This kind of image learning model based on deep learning has a very good generating capability, but usually requires a lot of training data and a long operation time, and the requirement of computing equipment is also expensive. For the general public, it usually depends on others to train a single-category generation model and is not possible to freely create multi-categories of images according to personal preferences.
The region clustering system proposed in this paper is intended to be applied to modular cartoon image creation We use the pre-trained convolutional neural network model to extract the features of input images’ regions, and then evaluating the cluster number of features by shallow network. At last, grouped these regions by unsupervised learning with the cluster number. Because of using shallow neural network, the computational cost and data volume requirements are lower compared to deep learning, and we don’t need any labels. By reducing the need for training data sets, the image generation system can more easily achieve multi-category image generation. The experimental results show that the system can automatically assess the number of better groupings and obtain good grouping results.

摘要 II
ABSTRACT VI
致謝 VII
目錄 VIII
圖目錄 X
表目錄 XII
第一章 緒論 1
1.1 研究動機 1
1.2 相關文獻 2
1.3 系統流程與論文架構 3
第二章 部位區域擷取 6
2.1 Statistical Region Merging（SRM） 6
2.2 去除輪廓線 9
2.3 連通分量標記 12
第三章 特徵擷取 17
3.1 卷積神經網路 17
3.1.1 卷積神經網路架構 17
3.1.2 VGG-16網路架構 22
3.1.3 Keras 25
3.2 主成分分析 25
3.3 Region Size與y軸位置資訊 28
第四章 部位分群 30
4.1 群數評估與分群運算 31
4.1.1 Simultaneous Two-Level – Self Organizing Maps 32
4.1.2 Kmeans++演算法 40
4.2 Clustering Refinement 44
4.3 聚類指數 50
第五章 實驗結果與討論 54
5.1 實驗環境與測試資料集 54
5.1.1 實驗設備 54
5.1.2 使用者介面 55
5.1.3 測試資料集 55
5.2 部位區域分群 56
5.2.1 S2L-SOM+Kmeans++之分群結果 56
5.2.2 不同群數之分群結果 57
5.2.3 群數評估 60
5.2.4 各群數之DBI比較 63
5.2.5 目標群數之分群結果準確率 64
5.2.6 S2L-SOM與S2L-SOM+Kmeans++之分群結果比較 66
5.2.7 搭配不同特徵之分群結果比較 67
5.3 系統運算時間 68
5.4 系統參數設置 69
5.4.1 不同維度特徵之分群結果比較 69
5.4.2 Clustering Refinement之y軸閥值比較 71
5.4.3 S2L-SOM之迭代次數比較 72
第六章 結論與未來工作 74
參考文獻 76

[1] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in neural information processing systems, pages 2672–2680, 2014.
[2] D. P. Kingma and M. Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013.
[3] Aaron van den Oord, Nal Kalchbrenner, Oriol Vinyals, Lasse Espeholt, Alex Graves, and Koray Kavukcuoglu. Conditional Image Generation with PixelCNN Decoders. arXiv preprint arXiv:1606.05328, 2016.
[4] Guilin Liu, Fitsum A. Reda, Kevin J. Shih, Ting-Chun Wang, Andrew Tao, and Bryan Catanzaro. Image Inpainting for Irregular Holes Using Partial Convolutions. arXiv preprint arXiv:1804.07723, 2018.
[5] Samaneh Azadi, Matthew Fisher, Vladimir Kim, Zhaowen Wang, Eli Shechtman, and Trevor Darrell. Multi-Content GAN for Few-Shot Font Style Transfer. arXiv preprint arXiv:1712.00516, 2017.
[6] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive Growing of GANs for Improved Quality, Stability, and Variation. arXiv preprint arXiv:1710.10196, 2018.
[7] Karol Gregor, Ivo Danihelka, Alex Graves, Danilo Jimenez Rezende, and Daan Wierstra. DRAW: A Recurrent Neural Network For Image Generation. arXiv preprint arXiv:1502.04623, 2015.
[8] Yanghua Jin, Jiakai Zhang, Minjun Li, Yingtao Tian, Huachun Zhu, and Zhihao Fang. Towards the Automatic Anime Characters Creation with Generative Adversarial Networks. arXiv preprint arXiv:1708.05509, 2017.
[9] Meng-Hang You. Automatic Cartoon Image Creation Through Learning from Examples. NCU CSIE, 2017.
[10] Nock, R. and F. Nielsen, Statistical region merging. IEEE Transactions on pattern analysis and machine intelligence, 2004. 26(11): p. 1452-1458.
[11] Samet, H.; Tamminen, M. (1988). "Efficient Component Labeling of Images of Arbitrary Dimension Represented by Linear Bintrees". IEEE Transactions on Pattern Analysis and Machine Intelligence.
[12] Michael B. Dillencourt; Hannan Samet; Markku Tamminen (1992). "A general approach to connected-component labeling for arbitrary image representations". Journal of the ACM.
[13] D. H. Hubel and T. N. Wiesel, “Receptive fields of single neurones in the cat’s striate cortex,”J. Physiol. (London) 148, 574–591 (1959).
[14] leonardblier. (2016, February 29). A BRIEF REPORT OF THE HEURITECH DEEP LEARNING MEETUP #5. from https://blog.heuritech.com/2016/02/29/a-brief-report-of-the-heuritech-deep-learning-meetup-5/
[15] Karen Simonyan, and Andrew Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556, 2015.
[16] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. ImageNet classification with deep convolutional neural networks. Proceeding：NIPS'12 Proceedings of the 25th International Conference on Neural Information Processing Systems - Volume 1, Pages 1097-1105, Lake Tahoe, Nevada, December 03 - 06, 2012.
[17] Pearson, K. On Lines and Planes of Closest Fit to Systems of Points in Space. Philosophical Magazine. 1901, 2 (6): 559–572.
[18] Guenael Cabanes, and Younes Bennani, "A simultaneous two-level clustering algorithm for automatic model selection.", IEEE International Conference on Image Processing, 2007.
[19] Kohonen, T. (1982). Self-organized formation of topologically correct feature maps. Biological Cybernetics, 43:59-69.
[20] MacQueen, J. B. (1967). Some Methods for classification and Analysis of Multivariate Observations. Proceedings of 5th Berkeley Symposium on Mathematical Statistics and Probability. 1. University of California Press. pp. 281–297.