Zero sample classification method based on antagonistic self-encoder model

Abstract

A zero-sample classification method based on antagonistic self-encoder model is provided. Using countermeasure self-encoder networks trained on visible classes, selecting network parameters w and v that best approximate the distribution of visual features and associate visual features with category semantic features, then the class semantic feature at of the unseen class is inputted into the network, and the Euclidean distance between the generated visual feature and the real visual feature is calculated by using the decoder network G. Finally, the category with the smallest distance is considered as the predicted category, so as to realize the task of zero-sample classification. The invention is more consistent with the characteristics of the real data, and simultaneously aligns the visual features and the category semantic features, so that better classification effect can be realized in the zero-sample task.

Description

technical field [0001] The invention relates to a zero-sample classification method. In particular, it relates to a zero-shot classification method based on an adversarial autoencoder model. Background technique [0002] Deep learning has greatly facilitated the development of computer vision, such as object classification, image retrieval, and action recognition. The performance of these tasks is usually evaluated after training with a large amount of labeled data. However, some tasks have only a small fraction of training data or even no training data, making traditional classification models perform poorly. In order to improve the classification performance of traditional classification models for classes with little or no data, zero-shot learning has attracted extensive attention. The task of Zero Shot Learning is to classify classes without training data. Humans have the ability to reason, which means that humans can successfully infer the category of unseen objects...

AI Technical Summary

Problems solved by technology

However, most adversarial networks only focus on generating distributions that approximate real visual features, but ignore the correspondence between visual features and category semantic features, making the generated visual features lack discriminative information to a certain extent.

Images