基于改进Resnet34的服装款式图廓形识别  

Garment pattern profile recognition based on improved Resnet34

作　　者：庹武[1] 刘永亮 高雅昆 郭鑫 魏新桥 杜聪 于媛媛[1] TUO Wu;LIU Yongliang;GAO Yakun;GUO Xin;WEI Xinqiao;DU Cong;YU Yuanyuan(College of Fashion,Zhongyuan University of Technology,Zhengzhou,Henan 451191,China;College of Electrical Engineering and Automation,Henan Institute of Technology,Xinxiang,Henan 453003,China)

机构地区：[1]中原工学院服装学院,河南郑州451191 [2]河南工学院电气工程与自动化学院,河南新乡453003

出　　处：《毛纺科技》2023年第6期95-102,共8页Wool Textile Journal

基　　金：河南省高等学校重点科研项目(19A540004,23A540007);河南省教育科学规划一般课题(2022YB0136)。

摘　　要：为解决服装款式图廓形识别技术复杂及识别精度不高的问题,提出一种基于残差神经网络ResNet34的改进方法。首先针对首层卷积提取图像的邻域信息范围较大的问题,优化首层网络结构提升其特征表达能力,然后针对残差块中线性变换和非线性变换不足的问题,引入融合非对称卷积和h-swish激活函数,并通过迁移学习提升模型的泛化能力;建立一个包括连衣裙、女上装和女裤3个大类共10种廓形的服装款式图数据集,对其数据增强后进行算法验证。实验结果表明:改进的ResNet34网络模型准确度达92.78%,相比ResNet34网络模型方法提升了3.2%,且性能优于ResNet50、MobileNetV2、CaffeNet等网络模型方法,可提高款式图廓形的识别精度,对服装企业智能制造过程中的款式识别具有借鉴意义。In order to solve the problems of complex recognition technology and low recognition accuracy,an improved method based on residual neural network ResNet34 was proposed.Firstly,to solve the problem of a large range of neighborhood information extracted by first-layer convolution,the first-layer network structure was optimized to improve its feature representation ability.Then,to solve the problem of insufficient linear and nonlinear transformations in residual blocks,fusion asymmetric convolution and h-swish activation function were introduced,and the generalization ability of the model was improved by transfer learning.A clothing style graph data set of 10 silhouettes in three categories including dress,blouse and pants was established,and the algorithm was verified after data enhancement.The experimental results show that:the accuracy of the improved ResNet34 network model is up to 92.78%,which is 3.2%higher than that of the ResNet34 network model method,and the performance is better than that of the ResNet50,MobileNetV2,CaffeNet and other network model methods,which can improve the accuracy of pattern profile recognition.It can be used as a reference for style recognition in intelligent manufacturing process of garment enterprises.

关 键 词：款式识别 服装款式 残差网络 激活函数 迁移学习 

分 类 号：TS941.26[轻工技术与工程—服装设计与工程]