fcanet

**FcaNet** (Frequency Channel Attention Network) is a **channel attention mechanism that replaces global average pooling with DCT (Discrete Cosine Transform) frequency components** — capturing richer channel statistics by using multiple frequency bases instead of just the DC (mean) component. **How Does FcaNet Work?** - **Key Insight**: Global average pooling = DC component of DCT. This captures only the mean and discards all frequency information. - **Multi-Frequency**: Use different DCT frequency components for different channel groups (low, mid, high frequencies). - **Channel Split**: Divide channels into groups, each processed with a different DCT basis. - **Attention**: Generate attention weights from the multi-frequency representation via FC + sigmoid. - **Paper**: Qin et al. (2021). **Why It Matters** - **Richer Statistics**: Captures frequency information beyond just the spatial mean (edges, textures, patterns). - **Drop-In**: Replaces GAP in any SE-style attention module with no architectural changes. - **Improvement**: Consistently outperforms SE-Net by using richer channel descriptors. **FcaNet** is **SE-Net with frequency vision** — replacing the simple mean pooling with multi-frequency DCT components for richer channel attention.