本研究發展一套使用共48盞RGB LED燈的影像擷取裝置，希望用此裝置來產生法向量貼圖(Normal map)。藉由電腦控制每一盞燈發光照射實驗材質時拍攝一張照片，根據所有拍攝照片合成法向量貼圖。本研究的影像擷取裝置採用標準球進行校正，實驗步驟為讀取照片中每一個像素的亮度數值，加上計算出的球體法向量，利用最小平方法迴歸得知每一盞燈的權重。最後根據使用每一盞燈的權重和拍攝照片每一個像素的亮度，即可製作出Normal map。
本研究製作出一張Normal map，最高可使用48張照片去製作，其中照射R、G與B的LED燈各16張。相較於目前市面上簡單即可以製作出Normal map的軟體，需花較多時間。我們嘗試把所需使用的照片數量減少製作Normal map，以降低所需時程。本研究嘗試使用24張(R、G與B各8張)及12張(R、G與B各4張)照片製作出Normal map，並和一般只使用一張材質照片推算出來的Normal map互相比較。由於Normal map屬於微小結構，並不容易透過直接取得表面的3D資料來驗證，故本研究採用人因實驗來評判法向量的優劣。我們主要選用布料為實驗驗證的材質項目，作為製作3D表面材質的範本。
本研究將人因實驗的結果做了ANOVA分析，結果顯示48盞、24盞與12盞燈所產生的Normal map並無明顯差異，意即12盞燈的配置即足夠產生可接受Normal map。與純軟體(Photoshop cc 2017)方法產生的Normal map相比，本研究所產生的Normal map貼圖效果普遍較好，但無統計上的顯著差異。因此本研究所發展的影像擷取裝置較適合顏色單一且不易反光的材質。

In this study, we develop a device which consists of 48 LEDs and a digital camera to capture Normal map. The device is able to individually control each LED and take photos for testing materials. Based on those taken photos under various lighting conditions, we synthesize then into a Normal map. Our equipment is calibrated using a standard ball. The experimental procedure is to read brightness of all pixels in images, and then, to determine a weighting function to fit the sphere equation by the least square method. Basis on the weights of LEDs, we finally can generate a Normal map.
In our experiment, we totally can make a Normal map from 48 photos, including 16 R, G and B LEDs. Comparing to the existing method, our method is time consuming. Therefore, we try to reduce the numbers of images and hope to have the similar Normal map quality. We have used 24 photos (R, G and B LEDs are 8 photos) and 12 photos (R, G and B LEDs are 4 photos) to make Normal maps. Then, we compare our methods with existing method. Since Normal map is very tiny structure, it is not easy to verify. Thus, we use subjective experiments to evaluate the quality of normal maps. We totally choose 9 materials which are mostly cotton material.
From the result of subjective experiment, we analysis the data by ANOVA. The result shows that there is no significant difference among the Normal maps from 48, 24 and 12 photos. That concludes that only 12 LEDs is enough to generate a quality Normal map. With comparing to the existing method, our result is slightly better. However, it is not significant different. And, our work is limited and only available for non-glossy materials so far.

[1]齋藤美穗‧矢野正‧五十嵐幹雄，“眼‧色‧光 優色再現求”， 社团法人日本印刷技術協会出版，第81-96頁，2007年。
[2]黃義淳，”3DS Max 2014 超強3D設計與絕佳動畫表現”，碁峰資訊股份有限公司出版，第8章，2013年。
[3]黃姿毓，“法向量貼圖製作臉部貼圖之探討” , 碩士論文，國立台灣科技大學，台北市，第11-16頁，2009年。
[4]楊雪果，”3DS Max 進階程序貼圖的藝術”，上奇資訊股份有限公司出版，2013年。
[5]V Krishnamurthy, and M Levoy, “Fitting smooth surfaces to dense polygon meshes,” ACM SIGGRAPH, 1996, pp. 313-324.
[6]J. Cohen, M. Olano, and D. Manocha, “Appearance-preserving simplification,” in Proceedings of the 25th Annual Conference on Computer Graphics and Interactive Techniques, 1998, pp. 115–122.
[7]P. Cignoni, C. Montani, C. Rocchini, and R. Scopigno, “A general method for preserving attribute values on simplified meshes,” IEEE Visualization '98. Proceedings, 1998, pp.59–66.
[8]C. T. Franklin, A. Finkelstein, and S. Rusinkiewicz, “Illustration of complex real-world objects using images with normals,” Non-photorealistic Animation and Rendering, 2007, pp.111–119.
[9]S. Ryu, S. H. Lee, and J. Park, “Real-time 3D surface modeling for image based relighting,” IEEE Transactions on Consumer Electronics, 2009, vol. 55, no. 4, pp. 2431–2435.
[10]M. K. Johnson, and E. H. Adelson, “Retrographic sensing for the measurement of surface texture and shape,” IEEE Computer Vision and Pattern Recognition, 2009, pp. 1070–1077.
[11]M. K. Johnson, F. Cole, A. Raj, and E. H. Adelson, “Microgeometry capture using an elastomeric sensor,” Advancing Computing as a Science & Profession Transactions on Geaphics, 2011, vol. 30, no. 4, pp.46:1–46:8.
[12]R. Pan, and V. Skala, “Normal map acquisition of nearly flatobjects using a flatbed scanner,” IEEE Virtual Reality and Visualization(ICVRV), 2013, pp.68–73.
[13]Y. Shi, B. Wen,W Ding, and N.Qi, “Prediction-based realistic 3D model compression,” Multimedia Tools and Applications, 2014, vol. 70, pp.2125–2137.
[14]B. Fodor, C. Kazó, Z. Janko, and L. Hajder, “Normal map recovery using bundle adjustment,” IET Computer Vision, 2014, pp.66–75.
[15]Y. Yamaura,T. Nanya, and H. Imoto, “Shape reconstruction from a Normal map in terms of uniform bi-quadratic B-spline surfaces,” Computer-Aided Design, 2015, vol. 63, pp.129–140.
[16]Autodesk 3DSMax overview, http://www.autodesk.com/products/3DS-max/overview (Accessed on Jan. 2017)
[17]ITU-R BT.500-13, “Methodology for the subjective assessment of the quality of television pictures,” 2012.
[18]張雲景‧張礽立‧賴礽仰，”物統計學－SPSS資料分析與研究設計概念”，高立出版，2008年。
[19]Z. Ji, X. Sun, S. Li, and Y. Wang, “Real-time bas-relief generation from depth-and-normal maps on GPU,” SGP 2014, pp. 75–83.
[20]O. Alexander, M. Rogers,W. Lambeth,M Chiang, and P.Debeve, “Creating a photoreal digital actor: the digital emily project,” Visual Media Production, 2009, pp. 176–187.
[21]H.Igehy, and L. Pereira, “Image replacement through texture synthesis,” IEEE International Conference on Image Processing, 1997, pp.186–189.
[22]M. Anderson, R.Motta, S.Chandrasekar, and M.Stokes, “Proposal for a standard default color space for the internet—sRGB,” Color and Imaging Conference, 4th Color and Imaging Conference Final Program and Proceedings, 1996, pp.238–245.
[23]李城中，”應用統計學：SPSS完全攻略”，新文京開發出版，第146-176頁，2008年。