Implicit semantic data enhancement method

Abstract

The invention discloses an implicit semantic data enhancement method. In the first stage, a feature extractor and a classifier are trained by using original data, then a knowledge graph is constructed by using the classifier, and a class center and a covariance matrix of each class are calculated based on extracted features; and in the second stage, similar categories of the tail categories are identified by using the constructed knowledge graph, and then the feature transformation direction of the similar categories is migrated to each tail sample for feature change. And the tail category diversity is greatly enriched by mining the correlation between the categories. And finally, carrying out infinite Gaussian distribution sampling on each sample along the migrated covariance matrix in the change direction to obtain infinite samples, and obtaining a new inference-based implicit semantic data enhancement loss by optimizing the upper bound of an infinite sample loss function to carry out model training.

Description

technical field [0001] The invention relates to a reasoning-based implicit semantic data enhancement method for solving long-tail classification, belonging to the technical field of computer software. Background technique [0002] With the rapid development of data acquisition technology, deep neural networks can achieve excellent performance on large-scale and evenly distributed training data. However, in real-world scenarios, the data is usually distributed with a long tail, that is, most categories belong to the tail category and only occupy a small number of samples, and a few head categories have most samples. In recent years, image classification for data with long-tail distribution is an important task in the image field. Effective use of unbalanced long-tail data to train good classifiers has received more and more attention. Existing methods generally improve the classification accuracy of tail categories by equalizing training data and data enhancement techniques....

AI Technical Summary

Problems solved by technology

However, the training of generative confrontation network is difficult, and as the number of enhanced samples increases, the model convergence speed drops sharply

Images