30 lines
No EOL
1.8 KiB
Markdown
30 lines
No EOL
1.8 KiB
Markdown
# mote-ocr
|
||
|
||
a ocr inspired by watamote, meant to be independent and free from bloat
|
||
|
||
<img src="https://files.catbox.moe/5zxbp7.gif" width="30%" />
|
||
|
||
### TODO
|
||
- add CTC decoder with dictionary and greedy search
|
||
- when finished , resume work on the cpp branch.
|
||
|
||
|
||
### it gets *most* of it correct.
|
||
| image | OCR result |
|
||
|-----------------------------------------------------------------------|--------------------------------------|
|
||
| <img src="https://git.ajattix.org/hashirama/mote-ocr/raw/branch/main/assets/examples/00.png" alt="Example 00" /> | 半刻前には見ず知らずの人 |
|
||
| <img src="https://git.ajattix.org/hashirama/mote-ocr/raw/branch/main/assets/examples/01.png" alt="Example 01" /> | 当民宿自慢の春の海鮮料理づくし！ |
|
||
| <img src="https://git.ajattix.org/hashirama/mote-ocr/raw/branch/main/assets/examples/02.png" alt="Example 02" /> | 樹とは本当の親子じやな |
|
||
|
||
## image processing </br><br>
|
||
as by now we use opencv, but we will migrate to a more self-contained (header-only) solution in the c++ version.</br><br>
|
||
for instance, this is how the preprocessing shoud look like: </br><br>
|
||
the resize is [done by ncnn](https://git.ajattix.org/hashirama/mote-ocr/src/commit/3609c5a6d6c642e16d73eca3d7ce8d881a86af4c/inference_test.py#L142) itself
|
||
| Before | After |
|
||
|--------|-------|
|
||
| <img src="https://git.ajattix.org/hashirama/mote-ocr/raw/branch/main/assets/examples/preprocessed/input.png" /> | <img src="https://git.ajattix.org/hashirama/mote-ocr/raw/branch/main/assets/examples/preprocessed/output.png" /> |
|
||
</br><br></br><br>
|
||
|
||
## mote-ocr in action:
|
||
|
||
<img src="https://git.ajattix.org/hashirama/mote-ocr/raw/branch/main/assets/2025-04-21_21-42_1.png" /> |