AI風格轉換圖像解析度優化研究之初探

A pilot study of Optimization for Image Resolution in AI Style Transfer

陳昭佑(Zhon-Yu Chen)；黃明俥(Ming-Jiu Huang)；闕家彬(Chia-Pin Chueh)

深度學習 ； 風格轉換 ； 影像處理 ； Deep Learning ； Style Transfer ； Image process

中華印刷科技年報

2022（2022 / 05 / 01）

123 - 131

繁體中文

隨著電腦硬體設備的大幅進步，人工智慧可應用在各領域中，其中在圖像設計方面，將圖像導入風格轉換成為人工智慧研究領域的重要主題之一，運用人工神經網路（ANN；Artificial Neural Network）所生成的風格轉換（Style Transfer）圖像，在數量和質量方面都優於其他應用；但是在印刷業圖像是會印刷在不同的材質上面，材質可能會是玻璃、紡織、塑膠…等，如果圖像的解析度不夠的話，就會造成印刷在產品端會出現圖像模糊的問題。本研究中將注重風格轉換後導致解析度降低的問題上，會使用即時超解析的方法解決這問題，利用特徵重建損失（Feature Reconstruction Loss）和風格重建損失（Style Reconstruction Loss）來調整圖像的解析度，讓圖像盡量接近原圖的解析度，且保留風格的紋路和顏色。另外，可以搭配自動化遮罩圖（Mask）生成的技術，主要功能是判斷原圖中顏色較深的用黑色取代，較淺的用白色取代，介於中間的用灰色取代，之後產生的遮罩圖可以應用在印刷的圖像處理上，使用在噴印堆疊和噴印出半透明、不透明的效果出來。最後的目標是希望經過風格轉換過後的圖，印刷在不同的材質上，能夠有高解析度的圖像品質呈現。

With the great advancement of computer hardware, artificial intelligence can be applied in various fields which include image design. The introduction of image style transferring has become one of the important themes in the field of artificial intelligence research. Using ANN (Artificial Neural Network) to produce images of style transferring is better than similar applications in other fields. However, in printing industry, images are printed on different materials, such as glass, textile, plastic, etc. If the resolution of images is not sufficient, the images will be blurred on products. In this study, the decrease of resolution after style transferring method is focused. So as to solve the problem, real-time super-resolution methods, such as feature reconstruction loss and style reconstruction loss, are applied to adjust the resolution of the images, so that the resolution of the style transferring images is as close to that of the original images as possible, and preserve the style texture and color. In addition, with the automatic mask generation technology, the technology is applied to judge the darker colors in the original images and to replace them with black, the lighter ones with white, and the intermediate ones with gray, and will then generate masks that can be applied to the image processing of printing, to produce printing stacks, and to print translucent or opaque effects. The final goal is to achieve high resolution image quality for printing on different materials after style transferring.

1. Final project: image quilting by ming hou.(2001)。取自 : http://vision.gel.ulaval.ca/~jflalonde/cours/4105/h14/tps/results/project/minghou/index.html
2. Building powerful image classification models using very little data.(2016)。 取自 : https://blog.keras.io/building-powerful-image-classification-models-using-very-little-data.html
3. Applied Deep Learning Convolutional Neural Networks. (2016)。 取自 : https://www.slideshare.net/ckmarkohchang/applied-deep-learning-1103-convolutional-neural-networks
4. Wikipedia. 四分位距 . (2021)。取自 : https://upload.wikimedia.org/wikipedia/commons/1/1a/Boxplot_vs_PDF.svg
5. Fast Neural Style-Transfer PyTorch.(2021)。取自 :https://github.com/matanby/fast-style-transfer-pytorch/tree/4d50bee0baf9e7f964ae6678e49cbd00de6a4a89
6. Wikipedia.Texture Synthesis.(2022)。 取自 : https://en.wikipedia.org/wiki/Texture_synthesis
7. Deng, Jia(2009).ImageNet: A large-scale hierarchical image database.2009 IEEE Conference on Computer Vision and Pattern Recognition
8. Gatys, Leon A.(2015).,未出版
9. Johnson, Justin(2016).,未出版
10. Ledig, Christian(2017).,未出版
11. Novak, Roman(2016).,未出版
12. Yann, LeCun(1998).Gradient-Based Learning Applied to Document Recognition.Proceedings of the IEEE,86(11)
13. Yeh James (2017)。卷積神經網絡介紹。技術文章，取自 : https://medium.com/jameslearningnote/資料分析-機器學習-第5-1講-卷積神經網絡介紹-convolutional-neural-network-4f8249d65d4f