Clothing retrieval technology based on deep metric learning

Abstract

The invention discloses a clothing retrieval technology based on deep metric learning, and relates to the field of content-based image retrieval. The method comprises the following steps: 1) model design: performing feature embedding on a batch of clothing pictures through a convolutional neural network and a full connection layer; 2) a sampling method: selecting a training sample picture sample pair according to cosine similarity, namely known similarity information of positive and negative sample pairs; and 3) loss function: substituting the picture sample pair selected by the sampling method in the step 2) into the designed loss function to calculate the loss, and performing back propagation to train the model. Image features are extracted by using the design of a convolutional neural network rear full connection layer, the performance is good and simple, and the parameters are few; according to the sampling method, information carried by simple and difficult sample pairs can be effectively mined, and meanwhile, unlimited punishment is prevented from damaging the overall structure of an embedded space. The loss function can adjust the loss of the difficult sample pair along withthe training process so as to fully learn the information of the simple sample pair and the difficult sample pair.

Description

technical field [0001] The invention relates to the field of content-based image retrieval, in particular to a clothing retrieval technology based on deep metric learning, which is mainly applied to the field of intelligent clothing item search on Internet e-commerce platforms. Background technique [0002] Apparel search is an important technology supporting the increasingly popular and generally prosperous e-commerce platform, which can greatly meet the needs of customers and promote the further popularization of e-commerce. There is an existing clothing image library, given a clothing image for query, the clothing retrieval task is to use the cosine similarity between the query image and the features extracted from the image in the library, and sort them according to the cosine similarity from large to small, with Obtain clothing pictures that are the same as or similar to the query picture in the picture library. [0003] Deep metric learning aims to learn an embedding ...

AI Technical Summary

Problems solved by technology

However, the importance of different image sample pairs is obviously different. If the weights are the same, the difficult sample pairs, that is, the important information, will be overwhelmed by the information of the simple sample pairs, which will affect the performance.

In the early stage of training, the network mainly learns the information of simple sample pairs. After mastering a large number of simple sample pairs, that is, in the middle and late stages, the network mainly learns the information of difficult samples, and some difficult sample information is difficult to grasp from the beginning to the end.

Images