本研究的目的為建立一套結合感性工學與類神經網路學習功能之汽車造形設計的輔助分析流程，以協助汽車設計師在構想發展階段時可快速獲得造形特徵的意象感覺，反之經由輸入意象感覺的感性語彙值也可以獲得相對應的汽車造形特徵，不僅增加汽車設計師創意構想，也可提昇設計的效率與水準。本研究搜集2001~2004年中的267輛車款正視圖為實驗樣本，進行類神經網路訓練與準確度評估，最後以五組不同向度的感性語彙值及2005年五種車款驗證已訓練完成的類神經網路，並評估其效益。

　　本研究選定「汽車正視圖」為實驗樣本，先將267輛樣本轉成灰階並且去背，然後進入Rhinoceros 3.0，運用控制點曲線(Control Point Curve)，將汽車樣本分為9個部分描繪成線稿圖。這9個部分包括：1.水箱護罩、2.標誌、3.中心引擎蓋、4.外部引擎蓋、5.擋泥版、6.引擎蓋延伸線、7.車頂輪廓線、8.車燈、9.汽車保險桿。接著運用Rhinoceros 3.0將9個部分線稿圖的點資料擷取出來並分別進行階層集群分析，再根據每個部份分群結果對267輛樣本進行編碼，作為類神經網路訓練的輸入層共33個神經元。另外，以20位具設計背景的受測者，配合5組感性語彙對將267輛汽車樣本線稿圖依據每組感性語彙對進行意象感覺的評分，即可獲得每款汽車樣本感性語彙的平均值，作為類神經網路訓練的輸出層共5個神經元。為了驗證類神經網路的訓練結果，本研究先以水箱護罩、車燈的分群結果，在同一群裡再進行K平均集群分析，找出群中心樣本20輛當作驗證用樣本，其餘247輛樣本作為類神經網路訓練之用。

　　經過倒傳遞類神經網路訓練九種類型中，觀察到以倒傳遞類神經網路V和VI的總平均誤差率最低 (均為13%) 為最佳訓練結果。最後輸入五組不同向度的感性語彙值，即可快速從資料庫中獲得相對應的建議車款。同時也可以輸入2005年五種車款，從已驗證訓練完成的類神經網路中，立即獲得相對應的汽車造形感性語彙值。未來汽車設計師便可以運用此汽車造形設計的輔助分析流程，在汽車造形設計的過程中，創造出更多、更有創意的汽車造形概念。

　This study aims at establishing a set of analytical process, which integrates kansei engineering and learning capability of neural network to assist vehicle designers in rapid vehicle styling design. Vehicle designers acquire feeling and ideas of styling features in conceptual development stage and, in turn, grasp corresponding vehicle styling features by inputting kansei word values of feeling. The process not only refines the original ideas of vehicle designers but also increases the efficiency and quality of styling design. This study collects the front views of 267 vehicle styles from year 2001 to year 2004 as experimental samples to conduct training and accuracy evaluation of the neural network. Five sets of kansei words are determined by four experts with design backgrounds and experiences. The values of the five sets of kansei words are collected using questionnaires.

　This study selects the front views of vehicles as experimental samples, converts the 267 samples to gray scale images, and then uses Rhinoceros 3.0 to draw the contours of the vehicle samples into nine parts of line arts using the Control Point Curve technique. The nine parts include (1) Grill, (2) Emblem, (3) Middle hood, (4) Outer hood, (5) Fender, (6) Hood flow line, (7) Roofline, (8) Head light, (9) Bumper. Next, point data of the nine parts of line arts are retrieved using Rhinoceros 3.0 for respective hierarchy cluster analysis and then 267 samples are encoded into 33 nerve cells as the input layer of neural network according to the grouping results of the cluster analysis. In addition, 20 subjects with design background are invited to rate the feeling of the line arts of the 267 vehicle samples based on the five pairs of kansei words respectively and obtain the mean value of the kansei word of each style of vehicle sample as output layer of the neural network.

　In order to verify the training accuracy and efficiency of the neural network, this study conducts a K mean grouping analysis using the grouping results of Grill and Head light, and obtains 20 group centers as verification samples and the remaining 247 samples as the neural network training samples. Through nine different types of back-propagation neural network training, total mean error rates of back-propagation neural network V and VI are the lowest (13%), which receive the best training results. Finally, five sets of kansei word values are input into the system and the system can quickly retrieve the corresponding recommended vehicle styles from the vehicle styling database. Meanwhile, five vehicle sample styles of year 2005 are input into the trained neural network to obtain the corresponding kansei word values. In the future, vehicle designers can use the analysis process of vehicle styling design to generate more original and more creative vehicle styling ideas in the conceptual design stage.

英文部分

[1] J. P. McCormacka, J. Cagan, and C. M. Vogel, “Speaking the Buick language: capturing, understanding, and exploring brand identity with shape grammars,” Design Studies, Vol. 25, Issue 1 , January 2004, pp. 1-29.

[2] M. Tovey and J. Owen, “Sketching and direct CAD modelling in automotive design,” Design Studies, Vol. 21, No. 6, November 2000.

[3] Y. Wang , B. Zhao, and L. Zhang, “Designing fair curves using monotone curvature pieces,” Computer Aided Geometric Design, Vol. 21, pp. 515–527, 2004.

[4] M. Tovey, “Styling and design: intuition and analysis in industrial design,” Design Studio, Vol. 18, No. 1, January, 1997.

[5] M. Tanaka, L. Anthony, and T. Kaneeda, “A single solution method for converting 2D assembly drawings to 3D part drawings,” Computer-Aided Design, Vol. 36, pp. 723–734, 2004.

[6] N. Allinson, H. Yin, and K. Obermayer, “Introduction: new developements in self-organizing maps,” Neural Networks, Vol. 15, 2002.

[7] S. Bandyopadhyay, “An automatic shape independent clustering technique,” Pattern Recognition, Vol. 37, pp. 33-45, 2004.

[8] N. Belacel, P. Hansen, and N. Mladenovic, “Fuzzy J-Means: a new heuristic for fuzzy clustering,” Pattern Recognition, Vol. 35, pp. 2193-2200, 2002.

[9] E. García-Berro, S. Torres, and J. Isern, “Using self-organizing maps to identify potential halo white dwarfs,” Neural Networks, Vol. 16, pp.405-410, 2003.

[10] N. Mistuo, “Kansei Engineering: A New Ergonomic Consumer-Oriented Technology for Product Development,” International Journal of Industrial Ergonomics, Vol. 15, pp. 3-11, 1995.

[11] L.J. Graham, K. Case, and R.L. Wood, “Genetic algorithms in computer-aided design,” Journal of Materials Processing Technology, Vol. 117, pp.216-221, 2001.

[12] M. Tovey, ”Concept design CAD for the automotive industry,” Journal of Engineering Design, Vol. 13, pp. 5-18, 2002.

[13] J.A.O. Simoes, “Icarus: the design process of a conceptual vehicle,” Materials and Design, Vol. 22, pp. 251-257, 2001.

[14] N. Efstratios, L. Luohui, L. Qi, “Neural networks and response surface polynomials for design of vehicle joints,” Computers and Structures, Vol. 75, pp.593-607, 2000.

[15] M. Shpitalni, H. Lipson, “3D conceptual design of sheet metal products by sketching,” Journal of Materials Processing Technology, Vol. 103, pp. 128-134, 2000.

[16] S. W. Hsiao, M.C. Liu, “A morphing method for shape generation and image prediction in product design,” Design Studies, Vol. 23, pp. 533–556, 2002.

[17] K. Sethi and I. Coman, “Image retrieval using hierarchical self-organizing feature maps,” Pattern Recognition, Vol. 20, Issues 11-13, pp. 1337-1345, 1999.

[18] K. Tsutsumi, “A development of the building Kansei information retrieval system,” International Conference on Computing in Civil and Building Engineering, Vol. 10, June 02-04, 2004.

[19] M. Nagamachi, “Kansei engineering: A new ergonomics consumer-oriented technology for product evelopment,” International Journal of Industrial Ergonomics, Vol. 15, pp. 3-10, 1995.

[20] Stephen Newbury, “The Car Design Yearbook,” Vol. 1, Merrell, 2002.

[21] Stephen Newbury, “The Car Design Yearbook,” Vol. 2, Merrell, 2003.

[22] Stephen Newbury, “The Car Design Yearbook,” Vol. 3, Merrell, 2004.

中文部分

[23] 邱文福著，大車拼─台灣車壇贏的策略，秀威資訊科技股份有限公司，2003。

[24] 約拿丹‧曼特爾著，蘇采禾譯，汽車大戰─全球汽車大廠的發跡、商機與趨勢，時報文化出版企業股份有限公司，1997。

[25] 周鵬程編著，類神經網路入門─活用Matlab，全華科技圖書股份有限公司，2004。

[26] 葉怡成編著，類神經網路模式應用與實作，儒林圖書有限公司，2003。

[27] 黃俊英著，多變量分析《第七版》，中國經濟企業研究所，2000。

[28] 陳晉玄，消費者對產品識別之視覺認知研究─以汽車造形為例，國立台北科技大學創新設計研究所，2003。

[29] 陳鴻源，汽車輪廓形態意象與區分特徵關係之研究，國立成功大學工業設計研究所，2001。

[30] 長町三生，感性工學和方法論，感性工學委員會，日本，1997。

[31] 蔡詩怡，汽車造形輪廓之型態特徵辨識與認知之研究，國立雲林科技大學工業設計研究所，2003。

[32] 李立群，系統化方法應用於電腦輔助車形建構模擬之研究，國立成功大學工業設計研究所，1995。

[33] 張華城，應用類神經網路模式於產品造形特徵辨識之研究，國立成功大學工業設計研究所，1999。

[34] 馬志朋，不同國別汽車造形意象研究，國立成功大學工業設計研究所，1995。

[35] 蔡宏政，汽車造形、色彩之人機模糊介面研究，國立成功大學工業設計研究所，1992。

[36] 張明勝，以認知及動態特性為基的互動式電腦輔助汽車造形設計模式研究，國立成功大學工業設計研究所，1994。

[37] 盧定乾，空間曲面分割法在汽車外形上之應用，國立成功大學工業設計研究所，1993。

[38] 翁嘉聲，汽車造形形變對於意象認知與美感反應之關係研究，國立台灣科技大學設計研究所，2003。

[39] 王鉅富，造形於形變過程中與情感意象之關係研究─以汽車造形為例，國立台灣科技大學設計研究所，2002。

[40] 梁德聰，立體物件形變之動態展示於多向度認知空間之應用─以汽車造形為例，國立台灣科技大學工程技術研究所設計學程，1999。

[41] 鄭忠杰，特徵參數化與網際網路於產品造形衍生之研究，國立成功大學工業設計研究所，2003。