Vol.1 No.1 2008

46/76

Research paper : Technologies for the design and retail service of well-fitting eyeglass frames (M. Mochimaru, et al.)−43−Synthesiology - English edition Vol.1 No.1 (2008) 3-D face and head shapes. Head data was analyzed by MDS, and virtual 3-D homologous face models with scores of +3 standard deviation (s.d.) or –3 s.d. for only one of the 4 scales, such as (+3 s.d., 0, 0, 0), (–3 s.d.,0, 0, 0), (0, +3 s. d., 0, 0), etc., were generated. Virtual shapes at the center of distribution (average shape) or located between 2 scales were also generated. In total, 18 representative face models were generated. The homologous model described in Section 3.1 did not contain color information, such as eyebrows and lips that influence impression; therefore, a 2-D frontal face model including such information (Figure 8) was generated for each of the virtual shapes. The 3-D model of an individual was deformed into each of 18 generated models using the Free Form Deformation technique. Using each of the obtained transformation grids, the scan data of the individual with color information was deformed. The obtained representative 3-D face shape with color information was then projected onto the frontal plane. From the projected 2-D face with color information, a 2-D frontal face model including eyebrows and lips was created. The average texture of the young adult male Japanese was mapped onto the 2-D face model [8]. Through this process, 18 representative face images with color information were generated. Twelve eyeglass frames of different materials, types and shapes were selected. In order to reduce the time needed for each experiment to 30 minutes, 12 sets of 18 images were selected from 216 images (18 faces × 12 frames). Every set contained different faces and different frames.Participants were 300 young adult females. They evaluated one set of images (18 images) randomly selected by the web questionnaire system. The impression rating was evaluated by the visual analog scale method (VAS). Participants evaluated the image by moving the pointer on the scale between each pair of words in the web system (Figure 9). A computational model to estimate the impression rating from shape factors of the face and the frame was obtained by multiple regression (stepwise procedure). The dependent variable was the impression rating for each pair of words, and explanatory variables were the face dimensions and proportions, eyeglass frame dimensions and proportions, principal components of lens shape descriptors. Multiple correlation coefficients ranged from 0.784 to 0.901 for all impression ratings. Dimensions and proportions of the faces and eyeglass frames, and principal components of lens shape were contained in the explanatory variables for all impression ratings (Table 1).For the validation study, the following 12 face images were used: (1) virtual face images that were not used in the main experiment to develop the model, (2) a virtual face image that was close to the average face, (3) deformed actual face images, and (4) images that were used in the main experiment. Twelve images of faces with eyeglass frames were presented to another 59 young adult female participants, and impression ratings of 7 pairs of adjectives for 12 images were obtained. Comparing the impression ratings for validation images of (4) with those for the same images in the main experiment, it was found that the participants of 0.8630.8710.8340.7840.8210.9010.892おおらかな─神経質な自然な─不自然なおしゃれな─ダサい若い─老けた明るい─暗い涼しい─暑苦しい優しい─こわい相関係数メガネ寸法顔寸法玉型PCA1131825811238562531911Word pair describing impressionRFace measurementEyeglass frame measurementPCs for lens shapeGentle-SternCool-HotCheerful-QuietYounger looking-Older lookingSophisticated-UnstylishNatural-UnnaturalOptimistic-NervousFig. 8 2-D front face model used. From Mukaida et al., 2002Fig. 9 Example of Web-based questionnaire for impression ratingTable 1. Number of variables adopted in estimation functions (stepwise multiple regression)

※このページを正しく表示するにはFlashPlayer9以上が必要です