Zero-shot image classification based on class representation learning and attribute embedding learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) classifies unseen classes using semantic information from seen classes.
• Current ZSL methods struggle with visual variations within attributes, causing noisy features and reduced performance.
• Limited labeled data is a key challenge addressed by ZSL.

Purpose of the Study:

• To propose a novel zero-shot image classification method, CRAE (Class Representation and Attribute Embedding).
• To enhance classification robustness and accuracy by combining class representation and attribute embedding learning.
• To address the issue of noisy attribute-level features in existing ZSL approaches.

Main Methods:

• Designed an adaptive softmax activation function to normalize attribute feature maps, reducing noise.
• Introduced attribute-level contrastive learning with hard sample selection to optimize the attribute embedding space.
• Incorporated class-level contrastive learning to improve feature separation between different classes.

Main Results:

• CRAE significantly outperforms existing state-of-the-art methods on benchmark datasets (CUB, SUN, AWA2).
• The adaptive softmax function effectively reduces noise and improves attribute feature discriminability.
• Attribute and class-level contrastive learning reinforce feature distinctiveness and class separation.

Conclusions:

• CRAE demonstrates superior capability in zero-shot image classification.
• The proposed method effectively handles visual variations within attributes, leading to more robust classification.
• CRAE offers a promising advancement for tackling the challenge of limited labeled data in machine learning.