摘要
随着服装电子商务的蓬勃发展,海量的服装图像数据被累积,对服装图像“以图搜图”成为了当前的一个热点研究方向。服装图像有着丰富的整体语义信息和大量细节信息,要对其实现精准检索是一项挑战性难题。传统的基于人工语义标注的服装图像方法和以人工设计的颜色与纹理等内容特征进行服装图像检索的方法均存在较大局限性。文中利用卷积神经网络多层特征融合提取特征,然后使用K-Means聚类加快服装图像的检索,充分利用深度卷积神经网络在图像特征提取上的有效性和层次性,融合不同卷积层次特征的细节信息和抽象语义信息以提升检索的准确度,并利用K-Means加快检索速度。所提方法首先对服装图像数据集进行统一的尺寸处理,然后利用卷积神经网络进行训练和特征提取,抽取出服装图像从低到高的多层次特征,进而将多种层次的特征进行融合,最终使用K-Means聚类方法对提取的图像库特征进行有效检索。在DeepFashion子类数据集Category and Attribute Prediction Benchmark和In-shop Clothes Retrieval Benchmark上的实验结果表明,所提方法能有效增强服装图像的特征表达能力,提高了检索准确率和检索速度,优于其他主流方法。
The booming of clothing e-commerce has accumulated a large amount of clothing image data,and the“image search”of clothing images has become a hot research direction.Apparel images have rich overall semantic information and a large amount of detailed information,and achieving accurate retrieval is a challenging problem.Traditional me-thods of clothing image based on artificial semantic annotation and methods of image retrieval based on artificially designed content features such as color and texture have significant limitations.This paper proposed a clothing image retrieval method based on multi-layer feature fusion and K-Means clustering of convolutional neural networks,which makes full use of the effectiveness and hierarchy of deep convolutional neural network in image feature extraction,fuses the detailed information and abstract semantic information of different convolutional hierarchical features to improve retrieval accuracy,and uses K-Means to improve the retrieval speed.The proposed method firstly performs uniform size processing on the clothing image data set,then uses the convolutional neural network for training and feature extraction,extracts multi-level features of the clothing image from low to high,and then fuses various levels of features.Finally,the K-Means clustering method is used to efficiently retrieve large-scale image data.The experimental results on the DeepFashion sub-category data set Category and Attribute Prediction Benchmark and In-shop Clothes Retrieval Benchmark show that the proposed method can effectively enhance the feature expression ability of clothing images,and improve its retrieval accuracy and retrieval speed.The proposed method is supprior to other mainstream methods.
作者
侯媛媛
何儒汉
李敏
陈佳
HOU Yuan-yuan;HE Ru-han;LI Min;CHEN Jia(Engineering Research Center of Hubei Province for Clothing Information,Wuhan Textile University,Wuhan 430200,China)
出处
《计算机科学》
CSCD
北大核心
2019年第B06期215-221,共7页
Computer Science
基金
国家自然科学基金面上项目(61170093)资助
