基于Faster R CNN的浅口高跟鞋款式识别

张飚雪 刘成霞

摘要： 目前网购平台的搜索对象仍以文字为主，致使搜索效率低下。文章以广受女性欢迎的浅口高跟鞋为例，对利用图片进行款式自动识别的技术进行探索。以网购平台上收集的3类浅口高跟鞋（粗跟、细跟、坡跟）共900张图片（每款随机抽取200张作为训练集，剩余100张作为测试集）为研究对象，然后利用深度学习中的Faster R CNN检测模型对浅口高跟鞋款式进行训练和测试识别。结果表明：无论以鞋跟为目标区域，还是以整只鞋为检测区域，利用该模型都能对浅口高跟鞋图像实现良好的款式识别，准确率可达94%以上，且不用经过人为特征提取，方便可行;Faster R CNN检测模型的总体精度和检测速度比R CNN、SPP-Net、FAST R CNN更优。

关键词： 深度学习;款式识别;浅口高跟鞋;Faster R CNN;卷积神经网络

中图分类号： TS941.2

文献标志码： A

文章编号： 10017003（2021）01007906

引用页码： 011112

DOI： 10.3969/j.issn.1001-7003.2021.01.012（篇序）

Style recognition of shallow opening high-heeled shoes based on Faster R CNN

ZHANG Biaoxue1， LIU Chengxia2

（1.Yuanpei College， Shaoxing University， Shaoxing 312000， China; 2.School of Fashion Design & Engineering，Zhejiang Sci-Tech University， Hangzhou 310018， China）

Abstract：

Nowadays， the search object of online shopping platform is still mainly text， thus leading to low search efficiency. This paper researched the technology on how to realize style recognition by images of shallow opening high-heeled shoes popular with female. A total of 900 pictures of 3 kinds of shallow opening high-heeled shoes（thick heel， thin heel and wedge heel） collected from online shopping platforms（200 pictures of each style were randomly drawn as the training set， and the remaining 100 pictures were used as the test set） were selected as the object of study. Then， Faster R CNN in deep learning was used to train and test to recognize the style of shallow opening high-heeled shoes. Results showed that， Faster R CNN could recognize the style well， whether the heel or the whole shoe was used for the object area， with the precision reaching 94%. And it was very convenient because of no need of manual feature extraction. Moreover， the whole accuracy and detection speed of Faster R CNN were better than that of R CNN， SPP-Net and FAST R CNN.

Key words：

deep learning; style recognition; shallow opening high-heeled shoes; Faster R CNN; convolution neural network

在電子商务已成必然趋势的当今社会，网购发展迅速，越来越多的人通过网购挑选服装服饰。然而目前线上购物查找大多还是使用文字检索，这一方式需要对产品进行文字标注，不仅要耗费大量人力，且文字描述的能力也很有限，因此买家要快速搜到满意的商品非常困难，尤其是对服饰这样的非标品更加如此。如果能实现图像检索，客户可以直接由图片搜到心仪的服饰[1]，无疑会大幅提高网购便捷性，提升用户购买欲。

数字图像技术的发展使其成为可能，目前已有较多的研究围绕服装款式的自动识别[2-4]展开，如利用服装局部HOG特征，结合关键尺寸进行款式分类[5];利用匹配卷积神经网络和超像素平滑等方法[6]进行语义分割，以及多人的服装分割算法[7]等。……