基于度量学习的服装图像分类和检索

摘要

On the problem of clothing image classification and retrieval, the general convolutional neural network has limited ability to identify because of diverse patterns and different backgrounds in image.To solve this problem, a convolution neural network method based on metric learning is proposed, in which the metric learning is based on the triplet loss, and the network has three inputs: the reference sample, the positive sample and the negative sample.By means of metric learning, it is possible to reduce the intra-class feature distance and increase the inter-class feature distance, so as to achieve the fine-grained classification.In addition, the images in different backgrounds are input into the training network as positive samples to improve the anti-interference ability.On the problem of clothing retrieval, a fine-grained retrieval method is proposed, which combines features of convolutional layers and fully-connected layers.The experimental results show that the introduction of metric learning can enhance the feature extraction ability of the network and improve the accuracy of classification, and the retrieval based on combined features can ensure the accuracy of the results.%在服装图像分类和检索问题上,由于服装花纹样式的多样性和图像中不同环境背景的影响,普通卷积神经网络的辨识能力有限.针对这种情况,提出一种基于度量学习的卷积神经网络方法,其中度量学习基于triplet loss实现,由此该网络有参考样本、正样本和负样本共三个输入.通过度量学习可以减小同类别特征间距,增大不同类别特征间距,从而达到细分类的目的.此外把不同背景环境下的图像作为正样本输入训练网络以提高抗干扰能力.在服装检索问题上,提出融合卷积层特征和全连接层特征的精细检索方法.实验结果表明,度量学习的引入可以增强网络的特征提取能力,提高分类准确性,而基于融合特征的检索可以保证结果的精确性.