.png){kind=icon}
.png){kind=icon}
Return
Attention-driven frequency-based Zero-Shot Learning with phase augmentation
.png){kind=icon}
PRE
AI
DOI：10.1007/s13042-024-02512-w.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Zero-Shot Learning (ZSL) aims to recognize unseen classes by aligning visual and semantic information. However, existing methods often struggle with noise in the RGB domain, which limits their ability to capture fine-grained semantic attributes, such as a grey bird's tail blending with the ground. This visual ambiguity in the RGB domain negatively impacts model performance. In contrast, the frequency domain can better capture high-frequency signals that are often overlooked in RGB, making areas that are easily confused in RGB more distinguishable. To address this issue, we propose a novel frequency-based framework that transforms spatial features into the frequency domain, allowing for more robust attribute representation and improved noise suppression. The framework incorporates a Multi-Scale Frequency Fusion Module that integrates multi-scale feature maps with frequency domain attention, and a Phase-based Augmentation Module that enhances key attributes by augmenting phase information. Additionally, we introduce two novel modules: the Masked Residual Aggregation Module for combining global and local features and the Phase High-Frequency Filtering Module for image denoising. The Mean Class Accuracy results of our method on CUB, AWA2, and aPY datasets are 2.8%, 5.0%, and 7.4% higher than other methods, respectively. We establish a new direction in frequency-based zero-shot learning. Source code at https://github.com/Waldeinsamkeit628/AFPA.
Keywords:
Zero-shot learning
Object recognition
Joint embedding
Journal
IF:
2.7
Papers:
3.1K
Citations:
5.6K