Training method, system, storage medium and computer equipment for trademark image retrieval model

Abstract

The invention relates to a trademark image retrieval model training method, comprising: obtaining multiple sets of sample data, selecting the most difficult positive sample and multiple difficult negative samples for each query sample according to the similarity; Difficult positive samples and corresponding multiple difficult negative samples are used as a set of training data, and the neural network is used to train the trademark image retrieval model according to multiple sets of training data; the trademark image retrieval model is updated according to the comparison loss function of multiple negative examples until If the verification effect of the trademark image retrieval model on the verification set is no longer improved, the training ends. The present invention eliminates easy samples and mines difficult samples according to the similarity, makes full use of a small number of difficult samples, and adjusts the parameters of the neural network in a more targeted manner, which can better delay model convergence / overfitting and make training more efficient. The fuller, the better. The invention also relates to a trademark image retrieval model training system, storage medium and computer equipment.

Description

technical field [0001] The invention relates to the technical field of model training, in particular to a trademark image retrieval model training method, system, storage medium and computer equipment. Background technique [0002] In the prior art, the training method of trademark image retrieval model generally adopts fixed or random positive and negative examples. Fixed positive and negative examples are likely to lead to model overfitting, that is, the effect is only good on the training data; random positive and negative examples are likely to lead to underfitting. Fitting and insufficient training, that is, most of the models seen are very simple cases, and there is no targeted training to improve error-prone cases. Contents of the invention [0003] The technical problem to be solved by the present invention is to provide a trademark image retrieval model training method, system, storage medium and computer equipment for the problems existing in the prior art. [0...

AI Technical Summary

Problems solved by technology

[0002] In the prior art, the training method of trademark image retrieval model generally adopts fixed or random positive and negative examples. Fixed positive and negative examples are likely to lead to model overfitting, that is, the effect is only good on the training data; random positive and negative examples are likely to lead to underfitting. Fitting and insufficient training, that is, most of the models seen are very simple cases, and there is no targeted training for error-prone cases to improve

Images