Generalized zero-sample image recognition method based on conditional adversarial automatic encoder

Abstract

The invention relates to a zero-sample image recognition method, in particular to a generalized zero-sample image recognition method based on a conditional adversarial automatic encoder. The inventiondiscloses the generalized zero sample image identification method based on a conditional adversarial automatic encoder, and the method employs a decoder in the conditional adversarial automatic coding machine to generate a pseudo sample through combining with prior distribution and class attribute conditions, and solves a zero sample problem. The invention has the beneficial effects that the sampling conditional adversarial automatic encoder matches integrated posteriori distribution of data potential semantic representation into prior distribution, such as standard normal distribution, so that the stability of a generation model is improved; class attributes are used as conditions to train an adversarial automatic coding machine, the discrimination capability of the model is improved, and generated pseudo samples are closer to real distribution of data; visible and invisible pseudo samples are generated at the same time, and only the pseudo samples are used to train the SVM classifier to eliminate the influence of data imbalance.

Description

technical field [0001] The invention relates to a zero-sample image recognition method, in particular to a generalized zero-sample image recognition method based on a conditional confrontation autoencoder. Background technique [0002] In recent years, with the explosive data growth on the Internet and new things emerging, it is difficult to collect enough labeled data to train image recognition models. For this reason, experts and scholars have proposed the concept of zero-sample learning, which has aroused extensive research interest. Zero-sample learning is inspired by the process of human cognition of new things, trying to recognize new things from limited knowledge or other known things, supplemented by corresponding knowledge descriptions. The visible class in zero-shot learning is the class that can be used in the training process, while the invisible class is the class that cannot be used in the training process and is used in the testing process. That is, knowledg...

AI Technical Summary

Problems solved by technology

The present invention can not only solve the zero-sample problem, but is also applicable to the few-sample problem

Images