结合注意力机制的跨域服装检索

针对跨域服装检索中服装商品图像拍摄严格约束光照、背景等条件,而用户图像源自复杂多变的日常生活场景,难以避免背景干扰以及视角、姿态引起的服装形变等问题.提出一种结合注意力机制的跨域服装检索方法.利用深度卷积神经网络为基础,引入注意力机制重新分配不同特征所占比重,增强表述服装图像的重要特征,抑制不重要特征;加入短连接模块融合局部重要特征和整幅图像的高层语义信息,提取更具判别力的特征描述子;联合分类损失函数和三元组损失共同约束网络训练过程,基于类别信息缩小检索范围.采用标准的top-k检索精度作为评价指标,选择DeepFashion数据集与当前跨域服装检索常用方法进行对比,文中方法在top-20检索精度对比中取得了最好的检索性能(0.503).实验结果表明,该方法能有效地处理视角、姿态引起的服装形变和复杂背景的干扰,同时不需要大量的样本标注信息,有效地提高了跨域服装检索的精度.

Cross-Domain Clothing Retrieval with Attention Model

To cope with the large discrepancy between online shopping images and street photos,where the former is taken in ideal conditions of good lighting,clean backgrounds while the latter is captured in uncontrollable conditions,we propose a cross-domain clothing retrieval network based on an attention model.The first is to introduce an attention model to redistribute the proportion of different features on the basis of deep convolutional neural networks,which enhance the significant features of clothing images and suppress the unimportant features maps.The second is to introduce a short connection module and combine with the attention feature maps and the convolutional feature maps to further generate discriminative feature vectors for retrieval.The classification loss function and triplet loss are introduced to constraint training process of network and reduce the retrieval scope based on category information.The standard top-k retrieval precision is adopted as the evaluation index.DeepFashion dataset is selected to compare with the current algorithm for cross-domain clothing retrieval.The algorithm obtained the best retrieval performance(0.503)in the comparison of top-20 retrieval accuracy.Extensive experiments have revealed that the proposed method can effectively deal with the distortion of clothing and complex background interference.Meanwhile,it does not need huge labeled samples.The accuracy of cross-domain clothing retrieval is also improved.