Upload 3 files

README (3).md

@@ -0,0 +1,75 @@
+---
+license: mit
+language:
+  - km
+library_name: onnx
+tags:
+  - ocr
+  - khmer
+  - text-detection
+  - text-recognition
+  - onnxruntime
+  - on-device
+pipeline_tag: image-to-text
+---
+
+# Khmer OCR (ONNX) — mirror for iAny
+
+On-device **Khmer optical character recognition**: a text **detector** + a text
+**recognizer**, both in ONNX, that read Khmer (and Latin) text out of photos and
+scanned pages — fully offline, in the browser (onnxruntime-web) and on mobile
+(onnxruntime-react-native).
+
+This repository is a **verbatim mirror** of the models from
+[**seanghay/KhmerOCR**](https://github.com/seanghay/KhmerOCR) (MIT), re-hosted so
+the [iAny](https://github.com/sengtha/iAny) app can download them from a stable
+URL. **The model weights are unmodified** — all credit for training goes to
+[Seanghay Yath](https://github.com/seanghay). Only the app's TypeScript
+pre/post-processing was re-implemented from the reference Python.
+
+## Files
+
+| File | Size | SHA-256 |
+| --- | --- | --- |
+| `det.onnx` | 10,867,112 B | `61fb440d2eecd561a4d5ccd41e3be5a14a882643f1ac0e1f05f9f53ff9098e67` |
+| `rec.onnx` | 13,801,555 B | `931e3c960815a6119731d63759eac2ef800318bf131b8a92fecdb580fbbf1e65` |
+
+## Models
+
+### `det.onnx` — text detector (YOLO-style, opset 12)
+- **Input:** `images` — `float32[1, 3, 1024, 1024]`, RGB, letterboxed (aspect
+  preserved, top-left, black padding), values `/255`.
+- **Output:** `output0` — `float32[1, 6, 21504]` → per anchor
+  `[cx, cy, w, h, score_class0, score_class1]`.
+- **Classes:** `0` = figure/image region, `1` = text line.
+- Post-process: score threshold, xywh→xyxy, NMS, divide by the letterbox ratio.
+
+### `rec.onnx` — text recognizer (CRNN + CTC, opset 14)
+- **Input:** `input` — `float32[batch, 1, 32, width]`, grayscale (ITU-R 601-2
+  luma), height 32, width proportional, values `/255`.
+- **Outputs:**
+  - `logits` — `float32[seq_len, batch, 102]` (CTC; class `0` = blank).
+  - `font_logits` — `float32[batch, 6]` (font-style classifier).
+- Greedy CTC decode: argmax per step, collapse repeats, drop blanks. Class `i`
+  maps to glyph `TOKENS[i - 3]` (classes `1`–`2` are special / unused).
+
+**Charset (`TOKENS`, 99 glyphs):**
+
+```
+កខគឃងចឆជឈញដឋឌឍណតថទធនបផពភមយរលវឝឞសហឡអឣឤឥឦឧឩឪឫឬឭឮឯឰឱឲឳាិីឹឺុូួើឿៀេែៃោៅំះៈ៉៊់៌៍៎៏័៑្។៕៖ៗ៘៛៝០១២៣៤៥៦៧៨៩៳
+```
+
+**Font styles (`font_logits`):** `Regular, Italic, Bold, BoldItalic, Moul, MoulLight`.
+
+## Intended use
+
+Offline Khmer document/photo OCR on low-end devices. Works best on reasonably
+clean, straight-on images with a real text layer of printed Khmer. Scanned/photo
+quality (skew, lighting, blur) affects accuracy; there is no OCR fallback for
+image-only content beyond these models.
+
+## License & attribution
+
+**MIT.** Original models and training © [Seanghay Yath](https://github.com/seanghay),
+from [seanghay/KhmerOCR](https://github.com/seanghay/KhmerOCR). Redistributed
+unmodified under the MIT License; this mirror claims no authorship of the weights.

det.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61fb440d2eecd561a4d5ccd41e3be5a14a882643f1ac0e1f05f9f53ff9098e67
+size 10867112

rec.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:931e3c960815a6119731d63759eac2ef800318bf131b8a92fecdb580fbbf1e65
+size 13801555