Denoised and Dynamic Alignment Enhancement for Zero-Shot Learning.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Zero-shot learning (ZSL) aims to recognize unseen categories by linking visual features with semantic information.
• Current methods struggle with crude semantic proxies and background noise, hindering optimal visual-semantic alignment.
• Refining visual features and semantic perception is crucial for advancing ZSL.

Purpose of the Study:

• To introduce a novel Alignment-Enhanced Network (AENet) for improved zero-shot learning.
• To enhance visual-semantic alignment by denoising visual features and dynamically generating semantic information.
• To overcome limitations of existing methods in capturing attribute variations and handling redundant backgrounds.

Main Methods:

• Developed a visual denoising encoder using a class-agnostic mask to filter redundant visual information.
• Introduced a dynamic semantic generator that adaptively creates content-aware semantic proxies guided by visual features.
• Integrated a cross-fusion module for comprehensive interaction between denoised visual features and dynamic semantic proxies.

Main Results:

• The proposed AENet effectively denoises visual features, making them adaptable to unseen classes.
• Dynamic semantic proxies generated by AENet capture fine-grained visual variations.
• Experiments across three datasets show AENet narrows the visual-semantic gap, setting a new benchmark.

Conclusions:

• AENet significantly enhances visual-semantic alignment in zero-shot learning.
• The method demonstrates superior performance in recognizing unseen categories by addressing limitations of prior approaches.
• AENet establishes a new state-of-the-art for zero-shot learning tasks.