|
|
--- |
|
|
license: mit |
|
|
library_name: pytorch |
|
|
pipeline_tag: image-segmentation |
|
|
tags: |
|
|
- document-image-restoration |
|
|
- mask-background-detection |
|
|
- preprocessing |
|
|
- dtsprompt |
|
|
- doctra |
|
|
model-index: |
|
|
- name: DocRes MBD (mbd.pkl) |
|
|
results: [] |
|
|
--- |
|
|
|
|
|
# DocRes MBD Weights (mbd.pkl) |
|
|
|
|
|
These are the **MBD weights** (`mbd.pkl`) used by the **DocRes** model (CVPR 2024), rehosted for use in the [Doctra](https://github.com/AdemBoukhris457/Doctra) library. |
|
|
|
|
|
--- |
|
|
|
|
|
## 📖 Source |
|
|
- Original repository: [ZZZHANG-jx/DocRes](https://github.com/ZZZHANG-jx/DocRes) |
|
|
- Paper: *DocRes: Dynamic Task-Specific Prompt for Generalist Document Image Restoration* (CVPR 2024) |
|
|
|
|
|
--- |
|
|
|
|
|
## ⚖️ License |
|
|
MIT License (see LICENSE file). |
|
|
Weights are redistributed under the same terms, with attribution to the original authors. |
|
|
|
|
|
--- |
|
|
|
|
|
## ✅ Intended Use |
|
|
The `mbd.pkl` weights are used for **Mask and Background Detection (MBD)**, a critical component of DocRes for: |
|
|
|
|
|
- Generating document masks |
|
|
- Producing background priors |
|
|
- Supporting the Dynamic Task-Specific Prompt (DTSPrompt) mechanism |
|
|
|
|
|
These weights are required to prepare task-specific prompts for the main `docres.pkl` model. |
|
|
|
|
|
--- |
|
|
|
|
|
## ⚠️ Limitations |
|
|
- Designed specifically for document mask/background detection. |
|
|
- Performance depends on the quality of scanned/photographed inputs. |
|
|
- Should be used in combination with the main DocRes weights (`docres.pkl`) for full restoration capability. |
|
|
|
|
|
--- |
