基于循环Faster R-CNN的衬衫领型精确识别  

Accurate recognition of shirt collar type based on recurrent Faster R-CNN

作　　者：张怡 侯珏 刘正[1,2,3] ZHANG Yi;HOU Jue;LIU Zheng(School of Fashion Design&Engineering,Zhejiang Sci-Tech University,Hangzhou 310018,China;Clothing Engineering Center of Zhejiang Province,Hangzhou 310018,China;Garment Digital Technology Zhejiang Engineering Laboratory,Hangzhou 310018,China)

机构地区：[1]浙江理工大学服装学院,浙江杭州310018 [2]浙江省服装工程技术研究中心,浙江杭州310018 [3]服装数字化技术浙江省工程实验室,浙江杭州310018

出　　处：《西安工程大学学报》2022年第4期10-18,共9页Journal of Xi’an Polytechnic University

基　　金：文化和旅游部重点实验室资助项目(19076223-B);嘉兴市重点研发项目(2021BZ10001)。

摘　　要：为了更好地识别服装细节要素特征,以衬衫领部为例,提出了一种循环结构的Faster R-CNN网络模型。采用Faster R-CNN网络完成领部的一级识别与其边界框的截取,再循环利用其网络组件实现领部二级细节特征的识别分类,采用权重惩罚方法设定了3组类权重,解决衬衫领部数据集各类别间样本不均衡问题。实验结果表明:改变类权重后网络的平均精度均值(mean average precision,mAP)由96.23%提高到98.79%,使用样本数量占比最高类加权系数来控制权重,能够有效提升模型的检测精度。一级识别后截取到的图像由于只包含领部区域,能够突出目标;二级分类可提高其分类性能,准确率达到97.09%。该方法进一步提高了领部细节特征的识别效果,为服装款式精确识别提供新的方法与思路。In order to better identify the detailed element features of garments,a Faster R-CNN network model with recurrent structure was proposed by taking the collar of shirt as an example.The Faster R-CNN network was used to complete the first-level recognition of the collar and the interception of its bounding box,and then the network components were recycled to achieve the recognition and classification of the second-level detail features of the collar.The weight penalty method was used to set three groups of class weights to solve the problem of imbalanced samples among each class in the shirt collar dataset.The experimental results showe that the mean average precision of the network increased from 96.23%to 98.79%after changing the class weights,and the weighting coefficient of the class with the highest percentage of samples to control the weights can effectively improve the detection precision of the model.Since the image intercepted after the first-level recognition only contained the collar area,which can highlight the target,the accuracy of the second-level classification reaches 97.09%,which can improve the performance.The method further improves the recognition effect of collar detail features,which provides a new method and idea for the accurate recognition of garment styles.

关 键 词：Faster R-CNN 衬衫领部 识别定位 细节特征 深度学习 

分 类 号：TS941.2[轻工技术与工程—服装设计与工程]