基于属性解纠缠表示的交互式服装图像检索  

Interactive Clothing Retrieval Based on Attribute Disentangled Representations

作　　者：黄晓菊 陈金广[1] HUANG Xiaoju;CHEN Jinguang(School of Computer Science,Xi'an Polytechnic University,Xi'an 710048)

机构地区：[1]西安工程大学计算机科学学院,西安710048

出　　处：《计算机与数字工程》2025年第3期829-834,共6页Computer & Digital Engineering

基　　金：陕西省重点研发计划项目(编号:2020GY-122);陕西省教育厅科研计划项目(编号:21JP049)资助。

摘　　要：交互式服装检索可以将查询图像中的特定属性替换为目标属性,以更新检索结果。然而属性学习嵌入空间存在语义纠缠,改变特定属性会导致其他属性发生改变,限制了检索结果的可控性和有效性。为了解决该问题,论文提出了一种属性保持解纠缠表示学习(APDRL)网络来实现服装的交互式检索。首先,该网络为服装图像每种属性类型学习特定的子空间,以获得属性分离表示;其次,利用联合损失函数优化网络模型;最后,根据内存块中的每个属性原型嵌入,操纵向量将查询图像中特定属性修改为目标属性。在DeepFashion数据集上测试,实验结果表明:APDRL方法在保留服装固定属性的同时还可以修改目标属性,具有较强的可控性和实用性,优于其他主流的检索算法。Interactive clothing retrieval can replace the specific attributes in the query image with the target attributes to update the retrieval search results.However,there are semantically entangled in the embedding space of attribute learning,changing a specific attribute will causes other attributes to change,which limits the controllability and effectiveness of the retrieval results.In order to address this problem,this paper proposes an Attribute Preserving Disentangled Representation Learning(APDRL)network to realize interactive clothing retrieval.Firstly,the network learns a specific subspace for each attribute type of clothing image to obtain the attribute separation representation.Then,the joint loss function is used to optimize the network model.Finally,the manipulation vector modifies the specific attribute in the query image to the target attribute according to each attribute prototype embedded in the memory block.Testing on the DeepFashion dataset,the results show that APDRL method modify the target attribute while retaining the fixed attribute of clothing,It has strong controllability and practicability,and is superior to other mainstream methods.

关 键 词：交互式服装检索 解纠缠表示学习 DeepFashion 目标属性 损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]