“低聚(低多邊形)藝術”是現代藝術和趨勢之一，通常與肖像畫相關聯，可以追溯到早期3D動畫的製作，其通過多邊形網格模擬3D對象。目前現有主要的兩種創建低多邊形肖像的方法其一為使用Blender、3DS Max、Photoshop或Adobe Illustrator等繪圖軟體，手動將原圖分割成多個純色多邊形，並將每個多邊形拼貼成一個整體。實務上，若不以人工生產此類圖像，我們需要對操作程序進行編程以使其自動化。另一種半自動化的程序因而產生，它藉由著名的基因演算法以Voronoi diagram的分割基礎上產生低多邊形圖像。然而，這種方法反而增加了製圖時間，即便只是製作一張圖都要耗費數小時甚至數日。為解決此困境，我們提出了一種新穎的人工智慧模型，該模型通過自我監督學習完成了這項任務並取得了壓倒性的性能。該模型受益於 ResNeXt 和 ASPP 組合的成熟特性。而繪製時間是我們的模型與其他方法的最大區別。當然，在我們的任務中，圖像的分割與聚類任務密切相關，探索數據集中的圖像中像素簇的相似性是合理的期待。實驗表明，純數據集使訓練損失在非常早期的時期迅速下降，而混合數據集使其始終緩慢下降並收斂到更高的損失。最後，我們將展示使用我們的模型產生的低聚藝術作品。

Low poly art is an art style related to minimalism and has become a trend. Dating back to the early days of 3D animation simulation that mesh objects through polygon, it is now considered a style. There are currently two existing methods for creating low poly portraits. One is to use drawing software such as Blender, 3DS Max, Photoshop, or Adobe Illustrator to manually divide the original image into multiple solid-color polygons, and shape each one into a whole. Instead of producing such images manually, we would need to program the operation to automate it. Another well-known genetic algorithm appears with semi-automatic help. However, it actually increases the production time. It takes hours or days on average to execute just one painting. To provide a remedy to this dilemma, we propose a novel artificial intelligence model with self-supervised learning that benefits from the mature properties of the combination of ResNeXt and ASPP, and outperforms the two existing methods in the drawing time. In our task, certainly, the partitioning of images is closely related to the clustering task. Exploring the similarity of clusters of pixels in an image is a pragmatic option in datasets. Experiments show that the pure dataset makes the training loss drop rapidly at very early epochs, while the mixed dataset makes it always drop slowly and converge to a higher loss. At last, we will exhibit the low poly art conducted with our model.

[1] Anthony J Madden (2016). Low-Poly Art: Relaxation Colouring Book for Adults. CreateSpace Independent Publishing Platform.

[2] C. Finney, Kenneth (2004). 3D Game Programming All in One. Cengage Learning PTR.

[3] Maddie Stearn (2017). The Simple Way to Create Low Poly Portraits in Photoshop. Retrieved from https://blog.storyblocks.com/tutorials/simple-way-create-low-poly-portraits-photoshop/

[4] Sebastian Proost (2020). Minimalist Art Using a Genetic Algorithm, a different take on Vermeer's Girl with a Pearl Earring. Retrieved from https://blog.4dcu.be/programming/2020/02/10/Genetic-Art-Algorithm-2.html

[5] M.Innat (2020). Wiki-Art: Visual Art Encyclopedia, A collection of online art home gallery. Retrieved from https://www.kaggle.com/ipythonx/wikiart-gangogh-creating-art-gan

[6] Franz Aurenhammer (2013). Voronoi Diagrams and Delaunay Triangulations (p. 7). World Scientific Publishing Company.

[7] Chai, T., Draxler, R. R. (2014). Root mean square error (RMSE) or mean absolute error (MAE)? – Arguments against avoiding RMSE in the literature, Geosci. Model Dev., 7, 1247–1250. https://doi.org/10.5194/gmd-7-1247-2014, 2014.

[8] Fernando A. Fardo, Victor H. Conforto, Francisco C. de Oliveira, Paulo S. A Formal Evaluation of PSNR as Quality Measurement Parameter for Image Segmentation Algorithms. arXiv:1605.07116. https://doi.org/10.48550/arXiv.1605.07116

[9] Jim Nilsson, Tomas Akenine-Möller. Understanding SSIM. arXiv:2006.13846. https://doi.org/10.48550/arXiv.2006.13846

[10] Doersch, Carl, Zisserman, Andrew (2017). Multi-task Self-Supervised Visual Learning. 2017 IEEE International Conference on Computer Vision (ICCV). IEEE: 2070–2079. arXiv:1708.07860. doi:10.1109/iccv.2017.226. ISBN 978-1-5386-1032-9. S2CID 473729.

[11] Dor Bank, Noam Koenigstein, Raja Giryes. Autoencoders. arXiv:2003.05991. https://doi.org/10.48550/arXiv.2003.05991

[12] Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, Kaiming He. Aggregated Residual Transformations for Deep Neural Networks. arXiv:1611.05431. https://doi.org/10.48550/arXiv.1611.05431

[13] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, Alan L. Yuille. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. arXiv:1606.00915. https://doi.org/10.48550/arXiv.1606.00915

[14] Diganta Misra. Mish: A Self Regularized Non-Monotonic Activation Function. arXiv:1908.08681. https://doi.org/10.48550/arXiv.1908.08681

[15] Thorndike, R.L. (1953). Who belongs in the family? Psychometrika, 18, 267-276.

[16] TensorFlow. Retrieved from https://www.tensorflow.org/