基于组合模型的车模识别算法研究
2021-11-28张一博茹禹然赵文吕王冬琦
张一博 茹禹然 赵文吕 王冬琦



摘要:针对小比例车模识别中图像种类繁多、部分类间相似度较高、网络数据类别不均衡以及质量参差不一的问题,文章提出了一种组合模型。首先对网络采集的图像数据设计了一种基于深度学习的方法进行清洗,然后以破坏-重建学习(Destruction and Construction Learning)方法为基础结合文章提出改进的Class-Balanced Focal Loss权重调节方法构建细粒度识别模型,最后文章选取了3种评价指标对模型效果进行评价。实验结果表明,该组合模型相较于原方法能更加准确地对小比例车模进行识别,对于少数类具备更强的泛化能力。
关键词: 深度学习; 小比例车模; 卷积神经网络; 细粒度识别; 不平衡数据集
中图分类号:TP311 文献标识码:A
文章编号:1009-3044(2021)28-0015-05
开放科学(资源服务)标识码(OSID):
Research on Car Model Recognition Algorithm Based on Combined Model
ZHANG Yi-bo,RU Yu-ran,ZHAO Wen-lv,WANG Dong-qi
(School of Software, Northeastern University, Shenyang 110000, China)
Abstract: In order to solve the problems of many kinds of images, high similarity among some classes, unbalance of network data classes and variable quality in small scale vehicle model recognition, a combined model is proposed in this paper. Image data from the first to the network design a deep Learning based method for cleaning, and then to Destruction and Construction Learning method combined with in this paper, on the basis of improving the Class-Balanced Focal Loss weight adjusting method to build fine-grained recognition model, at the end of the paper analyzes three kinds of evaluation index to assess the economic impact on the model. Experimental results show that compared with the original method, the combined model can identify the small proportion car models more accurately, and has stronger generalization ability for a few classes.
Key words:deep learning; small scale car model; convolutional neural network; fine-grained identification; imbalanced data set
所謂“小比例汽车模型”是指尺寸1:64或更小的汽车模型,因为小比例汽车模型同时具备可玩性和可收藏性的特点,所以长期以来一直受到孩子和收藏者的青睐。随着经济的发展和人民生活水平的日益提高,中国玩具市场消费进入高速增长阶段[1],国内正在形成自己的收藏文化,据不完全统计,国内各类贴吧和即时通讯群组活跃人数达到近百万且市场潜力巨大[2]。然而由于历史与文化原因,目前市场上流行的小比例车模以日本和美国品牌居多,这导致了消费者难以获取相关信息,同时利用信息差进行诈骗的行为也屡屡发生,这些都阻碍了车模市场的发展。为了解决这些问题,设计一个能快速准确识别车模类型的算法模型是有必要的。……
