Semi-Supervised Low-Rank Semantics Grouping for Zero-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot learning (ZSL) aims to classify unseen classes using knowledge from observed classes.
• Traditional ZSL relies heavily on semantic attributes, which are often impractical to annotate in real-world scenarios.
• Limited labeled data significantly hinders ZSL performance.

Purpose of the Study:

• To develop a semi-supervised method for zero-shot learning that overcomes the reliance on complete semantic attributes.
• To jointly learn visual-semantic relationships and recover missing label information for seen classes.
• To improve the discriminative ability of zero-shot learning models for unseen classes.

Main Methods:

• A Low-rank Semantics Grouping (LSG) method is proposed for semi-supervised zero-shot learning.
• A visual-semantic encoder is employed as a projection model.
• A low-rank semantic grouping scheme captures attribute correlations, and a Laplacian graph aids label propagation.

Main Results:

• The LSG method demonstrates significant efficiency compared to state-of-the-art methods on standard ZSL benchmarks.
• The model exhibits robustness across various levels of missing label data.
• Visualized results confirm LSG's enhanced ability to distinguish test unseen classes.

Conclusions:

• The proposed LSG method effectively addresses the challenge of incomplete semantic annotations in zero-shot learning.
• LSG improves the performance and robustness of zero-shot visual recognition systems.
• The approach facilitates more accurate classification of previously unobserved categories.