基于代价敏感的细粒度服装图片检索  

Fine-grained Apparel Image Retrieval Based on Cost-sensitive Method

作　　者：邹子安 何儒汉[1,2] ZOU Zian;HE Ruhan(School of Computer Science and Artificial Intelligence,Wuhan Textile University;Hubei Provincial Engineering Research Center for Intelligent Textile and Fashion,Wuhan 430200,China)

机构地区：[1]武汉纺织大学计算机与人工智能学院 [2]纺织服装智能化湖北省工程研究中心,湖北武汉430200

出　　处：《软件导刊》2023年第10期185-190,共6页Software Guide

基　　金：国家自然科学基金项目(61170093);湖北省教育厅科学技术研究计划重点项目(D20141603)。

摘　　要：细粒度服装图片类别间高度相似,在非平衡数据集下,其检索十分具有挑战性。为此,提出一种基于代价敏感的细粒度服装图片检索方法,通过设计一种代价敏感的损失函数针对性处理非平衡数据集问题,有效改进检索模型的关键点检测模块,改善细粒度服装图片检索性能。代价敏感损失函数策略基于固定加权和动态加权的类别再平衡,其中固定加权是基于类别频率和标签的预测概率赋值,而动态加权根据预测分数调整其权重,允许模型对不同难度的实例进行调整,从而提高困难类样本梯度更新权重。在服装数据集DeepFashion上的代价敏感损失函数消融实验结果表明,固定加权和动态加权均提升了非平衡数据集下的模型检索性能。与其他细粒度图像检索方法的比较实验进一步表明,代价敏感损失函数可以解决类别不平衡和困难类别的服装图片检索问题。此外,类别、风格等服装属性检索比较实验结果表明所提模型的改进和损失函数的优化策略有效。The high similarity between fine-grained clothing image categories makes their retrieval challenging in imbalanced datasets.There⁃fore,a cost sensitive fine-grained clothing image retrieval method is proposed.By designing a cost sensitive loss function to deal with unbal⁃anced data sets,the key point detection module of the retrieval model is effectively improved to improve the fine-grained clothing image re⁃trieval performance.The cost sensitive loss function strategy is based on the category rebalancing of fixed weighting and dynamic weighting,where the fixed weighting is based on the prediction probability assignment of category frequency and labels,while the dynamic weighting ad⁃justs its weight according to the prediction score,allowing the model to adjust instances of different difficulties,thus improving the gradient update weight of difficult class samples.The ablation experiment of the cost sensitive loss function on the clothing dataset DeepFashion shows that both fixed weighting and dynamic weighting improve the model retrieval performance under unbalanced data sets.Compared with other fine-grained image retrieval methods,the experiment further shows that the cost sensitive loss function can solve the problem of category im⁃balance and difficult category retrieval.In addition,the experimental results of clothing attribute retrieval such as category and style show that the improvement of the proposed model and the optimization of loss function are effective.

关 键 词：细粒度服装 图片检索 代价敏感 非平衡数据集 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]