Transductive Zero-Shot Learning With a Self-Training Dictionary Approach.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot learning (ZSL) addresses the challenge of recognizing object instances from unseen classes without direct training data.
• Existing ZSL methods focus on domain distribution connections and semantic interactions between image and label spaces.
• Domain shift remains a significant hurdle in achieving robust ZSL performance.

Purpose of the Study:

• To propose a novel bidirectional mapping-based semantic relationship modeling scheme for effective cross-modal knowledge transfer in ZSL.
• To address the domain shift problem inherent in ZSL.
• To enhance the object classification capacity for unseen classes.

Main Methods:

• A bidirectional mapping scheme projects image features and label embeddings into a common latent space, establishing cross-modal connections.
• A transductive learning approach refines class predictions iteratively using bootstrapping over reliable instances.
• The model progressively reinforces object classification capabilities through model updating.

Main Results:

• The proposed approach demonstrates superior effectiveness compared to state-of-the-art methods on four benchmark datasets.
• Experimental results validate the capability of bidirectional mapping for cross-modal knowledge transfer.
• The transductive learning strategy effectively mitigates the domain shift problem.

Conclusions:

• The bidirectional mapping scheme offers a powerful solution for modeling semantic relationships and enabling knowledge transfer in ZSL.
• The iterative refining process in transductive learning significantly improves classification accuracy for unseen classes.
• The proposed approach advances the state-of-the-art in zero-shot learning for computer vision applications.