Zero-sample image classification method and device based on double auto-encoders

Abstract

The invention discloses a zero-sample image classification method and device based on double auto-encoders, and relates to the technical field of image classification, visual and semantic features are projected to a public space to learn potential semantics, and a consistent weight matrix is constructed based on graph knowledge to enable double projections to keep a consistent data structure. An epsilon-traction technology is introduced, a visible class classifier based on label relaxation is designed, the discrimination of potential language meaning and the generalization ability of a model are enhanced, and the method comprises the following steps: acquiring a sample image; constructing a visual feature vector, then establishing visual and semantic feature spaces and constructing a consistency weight matrix, constructing a regularization self-encoder based on double graph embedding, introducing an epsilon-traction technology, and establishing a visible class potential semantic classifier based on label relaxation, training a double discrimination graph regularization self-encoding model to obtain a zero sample classification model, and obtaining class labels of unseen class test samples in a public space by using a distance calculation formula.

Description

technical field [0001] The invention relates to the technical field of image classification, in particular to a zero-sample image classification method and device based on a double autoencoder. Background technique [0002] Zero-shot classification is an important research direction in image classification and belongs to an important branch of transfer learning. Although researchers have proposed some solutions, the technology still faces many challenges. Among them are the semantic gap between the low-level visual features and the high-level semantics, and the lack of discrimination of semantic attributes. [0003] In recent years, image classification based on deep learning has made breakthrough progress, but collecting and labeling training images is a very time-consuming and laborious work. Therefore, some researchers proposed the concept of "zero-shot learning", that is, to identify unseen classes with missing labels by transferring the known class knowledge of labels....

AI Technical Summary

Problems solved by technology

However, this patent directly learns the visual-semantic projection from visual features to semantic features, and does not solve the semantic gap between the underlying visual features and high-level semantic features

Images