基于多尺度扩张注意力的Styleformer汽车造型设计  

Research on Automotive Styling Design Tools Based on Styleformer-M

作　　者：徐敏靖 范永胜 桑彬彬 苏谦 XU Minjing;FAN Yongsheng;SANG Binbin;SU Qian(School of Computer and Information Technology,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331

出　　处：《新疆大学学报（自然科学版中英文）》2024年第5期591-598,605,共9页Journal of Xinjiang University(Natural Science Edition in Chinese and English)

基　　金：国家自然科学基金青年科学基金项目“开放环境下多粒度稳健知识获取及增量学习方法研究”(62306054)。

摘　　要：为了提升汽车设计图像的质量,开发了两个专门的数据集并通过多尺度扩张注意力机制优化现有的图像生成模型,以便捕获汽车的主体特征和结构细节.首先,创建了“WhiteCarSet”和“WhiteCarContour”数据集,分别聚焦于汽车的本体特征采集和汽车轮廓的精确收集.其次,针对图像生成模型,我们对Styleformer进行了改进,引入了多尺度扩张注意力操作,以增强模型捕获长距离依赖和理解对象全局结构的能力.在多个数据集上的实验结果显示,改进后的模型FID值相较于传统模型实现了15.57%的提升.在相同的改进模型Styleformer-M而数据集不同的情况下,使用我们开发的数据集,FID值比其它数据集提高了约22.29%.In order to improve the quality of automotive design images,two specialized datasets were developed,and the existing image generation models were optimized through a multi-scale dilated attention mechanism to better capture the main features and structural details of cars.Firstly,the“WhiteCarSet”and“WhiteCarContour”datasets were created,focusing respectively on the collection of the intrinsic features of the car body and the precise collection of the car contours.Subsequently,the Styleformer image generation model was refined by introducing a multi-scale dilated attention mechanism,which bolstered the model’s capability to capture long-range dependencies and comprehend the global structure of objects.Ultimately,experimental results across multiple datasets demonstrated that the improved model achieved a significant enhancement in the Fr´echet Inception Distance(FID)score,with an increase of up to 15.57%compared to traditional models.Moreover,utilizing our custom-developed datasets under the same enhanced model,Styleformer-M,the FID scores improved by approximately 22.29%over other datasets.

关 键 词：汽车造型设计 Styleformer-M 智能设计 数据集 

分 类 号：TB472[一般工业技术—工业设计] TP391.41[自动化与计算机技术—计算机应用技术]