基于膨胀卷积残差网络的服装图像检索

CLOTHING IMAGE RETRIEVAL BASED ON DILATED CONVOLUTIONAL RESIDUAL NETWORKS

针对传统的特征提取方法不能有效提取服装图像的语义特征和相似度计算泛化能力差的问题,提出一种基于膨胀卷积残差网络(Dilated Convolutional Residual Networks,DCRN)的服装图像检索方法。将膨胀卷积大尺寸感受野的优势和残差网络提取语义特征的优势结合,有效提取服装图像的特征;提出一种混合距离度量算法(Mixed Distance measurement algorithm,MD),通过计算余弦距离和马氏距离之和进行度量学习,从而稳定高效地计算特征向量的空间距离。实验表明DCRN方法能有效提取服装浅层的细节信息和深层的语义信息;在服装检索上,DCRN+MD方法的准确率较FashionNet方法有明显提升。

Aiming at the problem that the traditional feature extraction methods can not effectively extract the semantic features of clothing images and the generalization ability of similarity calculation is poor,we propose a clothing image retrieval method based on dilated convolutional residual networks(DCRN).The advantages of large size receptive field of dilated convolution were combined with the advantages of the residual network in extracting semantic features to effectively extract features of clothing images.We proposed a mixed distance measurement algorithm(MD),which calculated the spatial distance of feature vectors stably and efficiently by calculating the sum of cosine distance and Mahalanobis distance.Experiments show that the DCRN method can effectively extract the shallow details and deep semantic information of clothing.In terms of clothing retrieval,the accuracy of the DCRN+MD method is significantly higher than that of the FashionNet method.