WaveSeg
Segmentation that listens to the high frequencies.
Upload a brain MRI (FLAIR) slice or a colonoscopy image, pick the matching domain, and WaveSeg (SegFormer-B0 + Frequency-Boundary Adapter gating) will segment it. Optionally upload a ground-truth mask to see the Dice score.
| Input image | Domain | Ground-truth mask (optional) |
|---|
WaveSeg adds a tiny (<0.2M-parameter) Frequency-Boundary Adapter (FBA) between a
SegFormer-B0 decoder and its segmentation head. FBA applies a Haar Discrete Wavelet
Transform to the decoder features, exposing their high-frequency sub-bands (LH/HL/HH) -
where object boundaries live - gates them through a 1x1 conv into a boundary attention
map, and uses that map to re-weight the features before the final prediction:
out = feat * (1 + alpha * attn), with alpha a learnable scalar.
A boundary-frequency auxiliary loss (explicitly supervising the attention map toward ground-truth edges) was also tested, at several weights - it did not improve results after seed-averaging on either dataset, so the shipped model (A1) uses FBA gating alone, trained end-to-end via the segmentation loss only. This negative result is reported honestly in MODEL_CARD.md rather than dropped.