Update README.md
Browse files
README.md
CHANGED
|
@@ -31,7 +31,7 @@ This model was created to close the gap between accurate text enhancement and th
|
|
| 31 |
- **Parameters:** 1,944,168 (~1.9 MB int8)
|
| 32 |
- **Input:** RGB text line strip, height 48 px; width multiple of 16 (pad before inference, crop after)
|
| 33 |
- **License:** Apache-2.0
|
| 34 |
-
|
| 35 |
### Training & Evaluation
|
| 36 |
Trained from scratch on 50,000 synthetic strips from the Versta Glyphmatte dataset (5 new strips per step × 20k steps) with Adam + OneCycleLR, dice/BCE losses on matte and weight plus colour regression losses. Exported to ONNX with inlined weights (fp32) and an fp16 half-precision variant; validated against the fp32 PyTorch reference across four strip widths (figure-of-merit: max |Δ| ≤ 2.2e-06) and evaluated against a fixed validation shard (per-pixel IoU for the matte, AUC for weight).
|
| 37 |
|
|
|
|
| 31 |
- **Parameters:** 1,944,168 (~1.9 MB int8)
|
| 32 |
- **Input:** RGB text line strip, height 48 px; width multiple of 16 (pad before inference, crop after)
|
| 33 |
- **License:** Apache-2.0
|
| 34 |
+
|
| 35 |
### Training & Evaluation
|
| 36 |
Trained from scratch on 50,000 synthetic strips from the Versta Glyphmatte dataset (5 new strips per step × 20k steps) with Adam + OneCycleLR, dice/BCE losses on matte and weight plus colour regression losses. Exported to ONNX with inlined weights (fp32) and an fp16 half-precision variant; validated against the fp32 PyTorch reference across four strip widths (figure-of-merit: max |Δ| ≤ 2.2e-06) and evaluated against a fixed validation shard (per-pixel IoU for the matte, AUC for weight).
|
| 37 |
|