Research map / Image generation / Image restoration
Fourier Series Expansion Based Filter Parametrization for Equivariant Convolutions
Qi Xie, Qian Zhao, Zongben Xu, Deyu Meng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use proposed Fourier-based bases instead of traditional 2D Fourier bases for filter parametrization | F-Conv · Filter parametrization 5x5 (0 and 45 deg simultaneously) | RMSE: 1.8e-01 -> 6.6e-02 | Improved |
| use proposed bases instead of harmonic bases | F-Conv · Filter parametrization 11x11 (0 and 45 deg simultaneously) | RMSE: 2.7e-01 -> 2.0e-02 | Improved |
| use proposed bases instead of 2D Fourier bases on random initial filters | F-Conv · Random filter parametrization 11x11 | RMSE: 9.5e-13 | No clear effect |
| use F-Conv instead of E2-CNN for equivariance error (trained) | F-Conv · Equivariance error DIV2K (trained) | RMSE: 0.18 -> 0.19 | No clear effect |
| use F-Conv instead of E2-CNN for equivariance error (random init) | F-Conv · Equivariance error DIV2K (random init) | RMSE: 0.30 -> 0.42 | Hurt |
| use F-Conv instead of plain CNN for equivariance error (random init) | F-Conv · Equivariance error DIV2K (random init) | RMSE: 1.43 -> 0.42 | Improved |
| use all 25 proposed bases instead of first 11 principal components (ablation) | F-Conv · MNIST-rot-12k (simple 6-layer network) | test error (%): 1.23 -> 1.13 | Improved |
| use F-Conv-reduce instead of E2-CNN at equal parameters | F-Conv-reduce · MNIST-rot-12k (simple 6-layer network) | test error (%): 1.23 -> 1.23 | No clear effect |
| use F-Conv instead of E2-CNN in larger MNIST-rot-12k network | F-Conv · MNIST-rot-12k (leaderboard comparison) | test error (%): 0.682 -> 0.671 | Improved |
| use F-Conv instead of CNN for EDSR x4 SR without augmentation | F-Conv · SR x4 (4 datasets avg, no aug) | PSNR (dB): 26.95 -> 27.15 | Improved |
| use E2-CNN instead of CNN in EDSR x2 SR | E2-CNN · SR x2 (4 datasets avg, no aug) | PSNR (dB): 32.39 -> 32.23 | Hurt |
| use PDO-eConv instead of CNN in EDSR x2 SR | PDO-eConv · SR x2 (4 datasets avg, no aug) | PSNR (dB): 32.39 -> 31.62 | Hurt |
Closest papers by meaning
- On the Fourier analysis in the SO(3) space : the EquiLoPO Network (ICLR 2025 · similarity 0.86)
- Rotation-Equivariant Self-Supervised Method in Image Denoising (CVPR 2025 · similarity 0.85)
- Relaxed Rotational Equivariance via G-Biases in Vision (AAAI 2025 · similarity 0.84)
- A Regularization-Guided Equivariant Approach for Image Restoration (CVPR 2025 · similarity 0.84)
- R2Det: Exploring Relaxed Rotation Equivariance in 2D Object Detection (ICLR 2025 · similarity 0.83)
- 3D Equivariant Pose Regression via Direct Wigner-D Harmonics Prediction (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
