Watermark Remover — Multi-Model v2
6 specialized UNet++ segmentation models for watermark detection, combined with pixel-wise max fusion.
Models
| File | Specialty | Encoder | Best Val IoU | Size |
|---|---|---|---|---|
| segmenter_centered_text.pth | centered_text | efficientnet-b4 | N/A | 84.0 MB |
| segmenter_line_pattern.pth | line_pattern | efficientnet-b4 | N/A | 84.0 MB |
| segmenter_logo.pth | logo | efficientnet-b4 | N/A | 84.0 MB |
| segmenter_overlay_text.pth | overlay_text | efficientnet-b4 | N/A | 84.0 MB |
| segmenter_repeated_text.pth | repeated_text | efficientnet-b4 | N/A | 84.0 MB |
| segmenter_tiny_corner.pth | tiny_corner | efficientnet-b4 | N/A | 84.0 MB |
Architecture
Image → [Model A, B, C, D, E, F] → pixel-wise max(masks) → LaMa inpainting → Clean image
Each model is trained on synthetic watermarks matching its specialty.
Fusion strategy: pixel_max (logical OR of all masks).
Usage
import segmentation_models_pytorch as smp
import torch
model = smp.UnetPlusPlus(encoder_name="efficientnet-b4", encoder_weights=None, in_channels=3, classes=1)
state_dict = torch.load("segmenter_repeated_text.pth", map_location="cpu")
model.load_state_dict(state_dict)
model.eval()
License
Apache-2.0
Author
DevynLabs — AI tools for creators.