Semi-Negative Contrastive Subclass Discriminative Network for Compositional Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Compositional Zero-Shot Learning (CZSL) aims to recognize images with known attribute-object pairs, reducing the need for extensive training data.
• Existing CZSL methods struggle with challenges like multiple attributes per object, data distribution imbalances, and representation discrepancies, hindering the identification of novel combinations.

Purpose of the Study:

• To address the limitations of current CZSL methods and improve the recognition of unseen attribute-object combinations.
• To develop a robust model capable of handling complex visual data and achieving higher accuracy in zero-shot learning scenarios.

Main Methods:

• Proposed a Semi-Negative Contrastive Subclass Discriminative Network (SN-CSDN) utilizing contrastive learning.
• Introduced a semi-negative sampling strategy to enhance inter-class discrimination and fine-grained subclass feature capture.
• Developed a decoupled network branch for improved attribute-object relationship modeling and compositional embedding generation, leveraging subclass information.

Main Results:

• The SN-CSDN method demonstrated significant performance improvements on three benchmark datasets.
• The semi-negative sampling strategy effectively improved the model's ability to distinguish between classes and recognize subtle variations.
• The decoupled network branch enhanced feature representation and mitigated sample imbalance issues, particularly in long-tailed distributions.

Conclusions:

• The proposed SN-CSDN method offers a reliable and effective solution for compositional zero-shot learning.
• The study highlights the importance of addressing data imbalances and improving feature representation for robust CZSL.
• The findings suggest a promising direction for future research in zero-shot learning and visual recognition.