Abstract
Generalized Zero-Shot Learning (GZSL) aims to recognize objects from both seen and unseen categories using shared semantic attributes. Existing methods often fail to account for the inherent asymmetry between global, abstract attributes and local, concrete ones, leading to sub-optimal modeling. We propose the Dual Expert Distillation Network (DEDN), a "divide-and-conquer" approach that employs two specialized experts: a Coarse Expert (cExp) for global features and a Fine Expert (fExp) for localized details. Our framework incorporates a Dual Attention Network (DAN) to capture both spatial and channel-wise information, alongside a Margin-Aware Loss (MAL) to balance seen and unseen class confidence. DEDN achieves state-of-the-art results on CUB, SUN, and AWA2 benchmarks, demonstrating significant improvements in attribute localization and classification accuracy.
Introduction
Zero-Shot Learning (ZSL) relies on shared semantic attributes (e.g., "has wings," "is yellow") to identify novel categories. The more challenging Generalized ZSL (GZSL) requires identifying both seen and unseen classes during inference. Current research faces two primary hurdles:
- Attribute Asymmetry: Models typically treat all attributes uniformly, despite differences between global (scenic) and local (beak shape) features. This creates modeling pressure and sub-optimal performance.
- Underutilized Channel Information: Most region-based methods focus on spatial attention but ignore the semantic richness stored in the channels of feature maps.
The DEDN Framework
The DEDN architecture addresses these gaps through a multi-expert system and enhanced attention mechanisms.
Dual Expert Architecture
DEDN splits attribute recognition into two specialized streams:
- Coarse Expert (cExp): A holistic model maintaining a global perception field across the entire attribute set.
- Fine Expert (fExp): A modular system of subnetworks, each focusing on specific attribute clusters (e.g., "head" vs. "torso"). These clusters can be defined manually or through automated methods like K-Means.
Dual Attention Network (DAN)
DEDN replaces standard backbones with the Dual Attention Network, which processes features through parallel Region-Attribute and Channel-Attribute branches. This ensures the model captures both where an attribute is and the semantic correlations within the feature channels.
Mutual Distillation and Margin-Aware Loss (MAL)
During training, cExp and fExp undergo mutual distillation to reach a consensus. To mitigate the bias toward "seen" classes, the Margin-Aware Loss (MAL) adaptively adjusts decision boundaries, ensuring "unseen" classes remain competitive during inference without sacrificing the accuracy of seen classes.
Evaluation and Results
DEDN was evaluated on three standard benchmarks: CUB (birds), SUN (scenes), and AWA2 (animals).
Key Results
- State-of-the-Art Performance: DEDN achieved an H-score of 70.4% on CUB and 43.5% on SUN, outperforming previous leading methods like VS-Boost and TF-VAEGAN.
- Attribute Localization: Heatmaps confirm that the Fine Expert (fExp) provides superior precision in locating small, local attributes compared to holistic models.
- Ablation Studies: Removing Channel Attention (CA) resulted in significant performance drops, highlighting its critical role in semantic feature extraction.
Conclusion
The Dual Expert Distillation Network demonstrates that a "divide-and-conquer" strategy effectively manages attribute asymmetry in GZSL. By combining specialized experts with dual-branch attention and margin-aware optimization, DEDN sets a new standard for robust attribute-based recognition. Future work will explore automated attribute clustering and scalability to larger datasets.
Summary updated