Discriminative and noise-robust embedding for zero-shot learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current zero-shot learning (ZSL) methods heavily depend on pre-trained networks for visual features.
• These features often contain distracting elements, reducing the effectiveness of knowledge transfer.
• Existing ZSL approaches overlook noise and irrelevant details in attribute representations.

Purpose of the Study:

• To propose a novel method, Discriminative and Noise-Robust Embedding (DNRE), for ZSL.
• To address limitations in visual feature extraction and attribute representation in ZSL.
• To improve the distinctiveness of critical information and the alignment between attributes and visual regions.

Main Methods:

• Introduced a Channel Covariance Adaptive Enhancement (CCAE) module to refine visual representations by emphasizing informative channels.
• Developed a Dynamic Calibration Mechanism (DCM) to suppress noise and irrelevant signals for better attribute-visual region alignment.
• Evaluated the proposed DNRE method on three widely-used ZSL benchmarks.

Main Results:

• The proposed DNRE method consistently outperformed state-of-the-art baselines across benchmarks.
• DNRE demonstrated superior performance, particularly in accurate and robust attribute localization.
• The CCAE module effectively captured higher-order dependencies and refined visual features.

Conclusions:

• DNRE significantly enhances knowledge transfer in zero-shot learning by improving feature distinctiveness and attribute alignment.
• The proposed modules (CCAE and DCM) effectively handle noisy attributes and distracting visual components.
• DNRE represents a robust and effective approach for advancing zero-shot learning capabilities.