一种服装轮廓提取及款式识别的方法  

Method for clothing outline extraction and style recognition

作　　者：胡新荣 欧阳祥 刘军平 HU Xinrong;OUYANG Xiang;LIU Junping(School of Mathematics and Computer,Wuhan Textile University,Wuhan 430000,China;Hubei Province Engineering Research Center for Clothing Information,Wuhan 430000,China)

机构地区：[1]武汉纺织大学数学与计算机学院,湖北武汉430000 [2]湖北省服装信息化工程技术研究中心,湖北武汉430000

出　　处：《现代电子技术》2021年第5期75-79,共5页Modern Electronics Technique

基　　金：国家自然科学基金项目(61103085)。

摘　　要：服装的款式识别是服装检索中的一个关键性问题。自然条件下的服装图像包含人体信息和复杂背景等,对服装款式特征提取产生干扰。针对以上问题,提出一种结合目标检测网络和GA-BP神经网络的服装款式识别方法。首先,利用目标检测网络Mask R⁃CNN对服装图像进行像素级分割,提取服装轮廓,并得到服装初步特征;随后针对处理后的服装轮廓,使用BP神经网络识别每种类别的服装轮廓,由于BP神经网络训练时易陷入局部最优,因此采用遗传算法优化的GA⁃BP神经网络,最后得到服装款式识别的结果。实验结果表明,与现有的算法相比,提出的算法识别率更高。Clothing style recognition is a key issue in clothing retrieval.Clothing images under natural conditions usually contain information of human body and complex backgrounds,which interferes with the feature extraction of clothing style.In view of this,a clothing style recognition method combining object detection network and GA⁃BP(genetic algorithm⁃back propagation)neural network is proposed.The object detection network Mask R⁃CNN(Region⁃CNN)is used to perform pixel⁃level segmentation on the clothing images,extract clothing outline,and obtain the preliminary clothing features.For the processed clothing outlines,the BP neural network is used to identify the clothing outlines of each category.Since the BP neural network is easy to fall into a local optimum,the GA⁃BP neural network optimized by genetic algorithm is adopted to finally obtain the results of clothing style recognition.Experimental results show that,in comparison with the existing algorithms,the proposed algorithm has a higher recognition rate.

关 键 词：服装款式识别 目标检测 像素分割 轮廓提取 网络优化 轮廓识别 

分 类 号：TN911.73-34[电子电信—通信与信息系统]