结合特征融合的跨域服装检索  

Cross-domain Garment Retrieval Combined with Feature Fusion

作　　者：魏雄[1] 乐鸿飞 余锦露 WEI Xiong;YUE Hongfei;YU Jinu(School of Computer and Artificial Intelligence,Wuhan Textile University,Wuhan 430200,China)

机构地区：[1]武汉纺织大学计算机与人工智能学院,湖北武汉430200

出　　处：《软件导刊》2023年第9期196-201,共6页Software Guide

摘　　要：跨域服装检索由于域间差异大难以准确检索,是一项具有挑战性的任务。现有基于卷积神经网络的跨域服装检索算法缺少对服装局部特征信息的利用,导致性能不佳。针对该问题,提出一种结合特征融合的跨域服装图像检索方法。该方法以深度卷积神经网络提取为基础,利用多尺度卷积和自我注意提取具有代表性的局部信息,利用Gem池化提取全局信息,并将局部信息与全局表示进行聚合,生成更适用于跨域图像检索的特征嵌入。同时采用三元损失、中心损失、分类损失、质心损失联合的损失函数约束训练过程,在检索阶段使用质心损失缩短检索时间。该方法在DeepFashion数据集中取得了良好的检索性能,top-50检索精度达0.864,与CTL方法相比提高了1.4%。实验结果表明,全局与局部特征融合的跨域服装检索方法能在保证较高检索效率的情况下有效提高检索精度。Cross domain clothing retrieval is a challenging task due to the large differences between domains,making it difficult to accurately retrieve.The existing cross domain garment retrieval algorithms based on convolutional neural network lack the use of local garment feature in-formation,resulting in poor performance.A cross domain clothing image retrieval method combining feature fusion is proposed to address this issue.Based on deep convolutional neural network extraction,this method uses multi-scale convolution and self attention to extract representa-tive local information,uses Gem pooling to extract global information,and aggregates local information with global representation to generate feature embedding more suitable for cross domain image retrieval.At the same time,the training process is constrained by the Loss function of ternary loss,center loss,classification loss and centroid loss,and the centroid loss is used in the retrieval phase to shorten the retrieval time.This method achieved good retrieval performance in the DeepFashion dataset,with a top-50 retrieval accuracy of 0.864,which is 1.4%higher than the CTL method.The cross domain clothing retrieval method that integrates global and local features can effectively improve retrieval ac-curacy while ensuring high retrieval efficiency.

关 键 词：服装检索 特征融合 跨场景 质心损失 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]