Zero-shot Image Classification Method Based on Regression Variational Autoencoder

Abstract

The present invention is based on a zero-sample image classification method of a regression variational autoencoder, and relates to a method for using electronic equipment to identify graphics. By training and aligning cross-reconstructed variational autoencoders, the regression network loss is calculated, and the overall model network is calculated. Loss L, train the classifier, calculate the classification accuracy rate, and complete the zero-sample image classification based on the regression variational autoencoder, which overcomes the lack of training samples in the existing technology of generalized zero-sample classification, and the generated samples lack semantic and Generate defects against which the network is prone to collapse.

Description

technical field [0001] The technical solution of the present invention relates to a method for identifying graphics using electronic equipment, specifically a zero-sample image classification method based on a regression variational autoencoder. Background technique [0002] With the continuous development of deep learning in the field of computer vision, the demand for training data is also expanding. However, in reality, it is difficult to obtain a large amount of labeled data for supervised training, so the study of zero-shot image classification problems is particularly important. The zero-sample image classification problem is to establish the connection between known class images and unknown class images through knowledge transfer, so as to realize the requirement that the model trained with known class images can classify images of unknown classes, so as to achieve zero-sample image classification. Sample image classification purposes. For example, for bears, zebras...

AI Technical Summary

Problems solved by technology

[0012] When solving the generalized zero-sample image classification problem, a part of the known class data set in the database is used as a training set, and the other part and an unknown class data set are used as a test set, and the data used include known class image features v seen Data and known class semantic features c seen data, and unknown class image features v unseen Data and Unknown Class Semantic Features c unseen Data, all data are from the selected data disclosed in the document "Zero-Shot Learning-A Comprehensive Evaluation of the Good, the Bad and the Ugly", all image features are from the well-known residual network ResNet-101 of 2048 Dimensional final pooling layer, the data used is manually marked with attribute information as semantic features to complete data preprocessing;

Images