Training and Feature Extraction Method of Feature Extraction Network Based on Multiple Datasets

Abstract

The invention relates to the technical field of image processing, in particular to the training of a feature extraction network based on multiple data sets and a feature extraction method. The training method includes acquiring multiple sample data sets; sequentially inputting sample images in any two sample data sets into the feature extraction network , obtain the features corresponding to the sample images; perform gradient inversion on the features corresponding to all sample images to obtain the gradient inversion processing results corresponding to the sample images; based on the gradient inversion processing results, determine the difference between any two sample data sets The distance of the bulldozer; according to the distance of the bulldozer, the feature extraction network is trained to obtain the target feature extraction network. By calculating the bulldozer distance between any two sample data sets on the basis of the gradient reversal processing results, since the bulldozer distance can still reflect the distance between the two distributions when the two distributions do not overlap, it can be To some extent, the gradient dispersion phenomenon in GRL is alleviated.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to the training and feature extraction method of a feature extraction network based on multiple data sets. Background technique [0002] The rise of deep learning technology has greatly promoted the development of large-scale image retrieval. Deep learning has a great demand for data. The richer the data, the better the results. However, data labeling requires a lot of manpower and time. Joint training of datasets becomes a more convenient way to increase the size of the data. [0003] However, since the backgrounds and application scenarios of different datasets are different, joint training of multiple datasets will bring about the problem of domain offset and bring noise to feature learning. Some existing methods for solving the domain shift problem are called domain transfer. For example, domain transfer is realized based on gradient reverse layer (GRL), and most of th...

AI Technical Summary

Problems solved by technology

[0005] In view of this, the embodiment of the present invention provides a method for training and feature extraction of a feature extraction network based on multiple data sets to solve the problem of gradient dispersion in GRL training

Images