基于可变形卷积的服装检索方法

A garment retrieval method based on deformable convolution

传统的服装检索方法使用固定形状的感受野,当服装目标存在几何变形时无法有效地提取其特征。针对这个问题,提出基于可变形卷积和相似性学习的服装检索方法。首先,构建可变形卷积网络,自动学习服装特征的采样位置和服装图像的哈希编码;然后,级联相似性学习网络,度量哈希编码的相似性;最后,根据相似性评分产生检索结果。实验结果表明,该方法能够有效地提取存在几何变形的服装目标的特征,从而减少了图像背景特征的干扰,提高了检索模型的准确率。

Traditional garment retrieval methods use fixed-shape receptive fields, and they cannot extract features effectively when the garment target has geometric deformation. To solve this problem, we propose a garment retrieval method based on deformable convolution and similarity learning. Firstly, we build a deformable convolutional network which can automatically learn the sampling locations of garment features and the Hash code of garment images. Secondly, a similarity learning network is cascaded to measure the similarity of the Hash code. Finally, we obtain the retrieval results according to similarity scores. Experimental results show that this method can effectively extract the features of garment objects with geometric deformation, thus reducing the impact of image background features and improving the accuracy of the retrieval model.