Abstract: To solve the problem of insufficient confidence and real-time evaluation in the automotive appearance design process,a large data-driven automatic evaluation method for automotive appearance design was presented.From the perspective of consumers and styling experts,the large-scale car appearance multi-view images and the corresponding ratings,comments and car brand information were collected and collated.The rating information was processed by a comprehensive weighted scoring mechanism which considered the confidence level of the consumers.A large-scale automotive appearance evaluation dataset (DATE2022) including automotive appearance rating dataset of consumers and styling experts and the semantic evaluation dataset was created by filtering the commentary information using the NLP extraction method.The consumer and expert automatic scoring evaluation machines were obtained through deep learning regression method respectively.The perceptual image evaluation dataset of users and experts was classified by multi-label classification method to obtain two perceptual image evaluation machines of users and experts as well.The feature visualization method was used to locate the characteristic features of automobile appearance.Numerical experiments showed that the proposed method provided automatic prediction for automotive appearance design from consumers and experts,such as quantitative score,brand genes analysis and semantic perceptual information.

Key words: automobile design, deep learning, quantitative evaluation, evaluation confidence, preference prediction

摘要： 针对汽车外观设计过程中评价置信度与实时性不足的问题,提出一种大数据驱动的汽车外观设计评价方法。收集并整理了大规模汽车外观多视角图片数据及基于用户、专家视角所对应的评分、评论信息,对评分信息采用考虑评分置信度的综合加权评分机制进行处理;筛选评论信息采用感性意象词汇提取方法,用评分、感性意象词汇对相应的车型图片进行标注,创建了汽车外观评价数据集(DATE2022),数据集包括基于用户及专家视角的汽车外观评分数据集和感性意象评价数据集。对用户、专家评分数据集采用深度学习回归方法得到用户及专家两个自动评分机,对用户、专家感性意象评价数据集采用多标签分类方法得到用户及专家两个感性意象评价机,采用特征可视化方法定位用户、专家感兴趣的外观显著性特征区域。数值实验表明,该方法可以对汽车外观设计提供用户和专家两方面的量化评分、品牌基因、感性意象等自动预测。

关键词: 汽车设计, 深度学习, 量化评价, 评价置信度, 感性意象