Phase-Aware Scattering Encoder-Decoder for Dense Predictions
A research paper proposes Phase-Aware Scattering Encoder-Decoder (PASED), a novel architecture that preserves phase information in skip connections to restore spatial structure lost in scattering transforms. On BSD68 image denoising, breaking translation invariance improves PSNR by +2.17 dB, and phase preservation adds +1.03 dB. A spatial shuffling ablation results in a -1.26 dB penalty, demonstrating that phase encodes location-dependent structure. Preliminary extensibility is shown on ISIC skin lesion segmentation, with full cross-validation ongoing. The work advances principled wavelet-deep learning integration, addressing the stability-expressiveness trade-off in pixel-level prediction.
Key facts
- Proposes Phase-Aware Scattering Encoder-Decoder (PASED) for dense predictions
- Preserves phase in skip connections to restore spatial structure
- On BSD68 denoising: breaking translation invariance improves PSNR by +2.17 dB
- Phase preservation adds +1.03 dB on BSD68
- Spatial shuffling ablation yields -1.26 dB penalty
- Preliminary extensibility on ISIC skin lesion segmentation
- Full cross-validation for ISIC is ongoing work
- Advances wavelet-deep learning integration
Entities
—