keyword
frequency-aware transformer
A frequency-aware transformer is a neural network architecture that explicitly analyzes, separates, and processes multi-scale and directional frequency components within visual data using transformer-based attention mechanisms. Unlike standard vision transformers that process spatial patches uniformly, a frequency-aware transformer incorporates specialized attention and modulation modules to decompose signals into distinct frequency bands, distinguishing between broad structural components and fine, high-frequency directional textures. By applying tailored self-attention across these separate frequency representations and adaptively modulating them, the architecture reduces representational redundancy, captures anisotropic spatial patterns, and preserves directional details for tasks such as learned data compression and image restoration.
1 item
