| --- |
| license: apache-2.0 |
| language: |
| - en |
| pipeline_tag: image-to-text |
| tags: |
| - ocr |
| - onnx |
| - paddleocr |
| - rapidocr |
| - cpu |
| - quantized |
| base_model: |
| - PaddlePaddle/PP-OCRv6_tiny_det_onnx |
| - PaddlePaddle/PP-OCRv6_tiny_rec_onnx |
| base_model_relation: quantized |
| datasets: |
| - jsdnrs/ICDAR2019-SROIE |
| model-index: |
| - name: TinyOCR v6-w8-rec |
| results: [] |
| --- |
| |
| # TinyOCR 0.1.0 |
|
|
| English-first, offline CPU OCR with **3.10 MB of ONNX model files**. This is a |
| compressed pretrained PaddleOCR PP-OCRv6 model pair, not a newly trained TinyOCR |
| student. The detector is unchanged; the recognizer stores selected weights as |
| per-channel int8 values with constant decoding to float32. Computation remains |
| float32. No training or calibration was performed. The dataset tag describes |
| evaluation, not TinyOCR training. |
|
|
| ## Use |
|
|
| Python 3.12: |
|
|
| ```sh |
| pip install tinyocr-onnx==0.1.0 |
| tinyocr receipt.png |
| tinyocr receipt.png --json |
| ``` |
|
|
| ```python |
| from tinyocr import TinyOCR |
| |
| ocr = TinyOCR() |
| result = ocr("receipt.png") |
| print(result.text) |
| print(result.lines) |
| print(result.words) |
| ``` |
|
|
| The wheel bundles these exact model files and verifies their SHA-256 hashes. |
| It needs no model download or HF account after installation. Dependencies and |
| runtime memory are larger than the model files. Inputs are local image paths; |
| outputs contain text, line polygons/xyxy bounds and native word polygons in |
| original-image pixels. See the [runtime documentation](https://github.com/MMWilliams/tinyocr/tree/master/python). |
|
|
| ## Artifacts and provenance |
|
|
| | File | Bytes | SHA-256 | |
| | --- | ---: | --- | |
| | `PP-OCRv6_det_tiny.onnx` | 1,829,618 | `f42c0fbd294d95eac1a550e131b277dac97462c8025fa4b6c3cec1b7894bd3d5` | |
| | `PP-OCRv6_rec_tiny.onnx` | 1,269,595 | `38cd559cb4b24c5056592d981b13e086f9472db9b3c8968296f0d0c2b4a65531` | |
|
|
| Total: 3,099,213 bytes, versus 6,319,431 original bytes (50.957% smaller). |
| Original recognizer SHA-256: `e16e242de5937ad92609223f19bc2aff3727ee40b095f996907c24749bad251b`. |
| Source ONNX files came from RapidAI/RapidOCR v3.10.0 on ModelScope. NOTICE gives |
| the exact URLs and upstream attribution. `export-provenance.json` records |
| modified tensors; the export code is in the |
| [source repository](https://github.com/MMWilliams/tinyocr). |
|
|
| ## Measured results |
|
|
| 5,089 images per model with 120 prior-parent exclusions. These are comparative |
| development results; public pretraining overlap is unknown. |
|
|
| | Dataset | Images | Original E2E F1 (%) | TinyOCR E2E F1 (%) | Candidate p50 change | |
| | --- | ---: | ---: | ---: | ---: | |
| | SROIE | 341 | 41.036 | 40.779 | +21.23% | |
| | TextOCR | 3,027 | 43.412 | 43.334 | -6.91% | |
| | HierText | 1,721 | 40.330 | 40.204 | +16.16% | |
|
|
| Small observed E2E F1 losses of 0.079–0.257 percentage points, not exact accuracy |
| preservation. SROIE/TextOCR use the shared HierText polygon evaluator, not their |
| native leaderboard protocols. HierText uses its native word evaluator on the |
| overlap-filtered subset. Latency is one pass per model on the same Windows CPU |
| environment, not a repeated speed study. Peak process working set was |
| 498.84 MiB original / 500.10 MiB candidate. **No general speedup or RAM saving is |
| claimed.** See the [full benchmark](https://github.com/MMWilliams/tinyocr/blob/master/BENCHMARK_RESULTS.md) |
| for counts, source rights, protocols, timing and caveats. Images/labels are not |
| included in this model repository. |
|
|
| ## Configuration and limitations |
|
|
| RapidOCR 3.10.0; ONNX Runtime 1.31.0; one-thread CPU sessions and OpenCV; detector |
| ImageNet normalization, maximum side 960, threshold 0.2, box threshold 0.4, |
| unclip 1.4, no dilation; recognizer batch one; no orientation classifier. |
| The installed API matches the frozen research text/line/word adapter on three |
| development fixtures. English-first evaluation does not establish multilingual |
| performance. Rotations, difficult scenes and complex layouts can fail. |
| Reading order is geometric, not layout understanding. Native word polygons can |
| have zero area; they are retained for benchmark parity. No mobile performance |
| claim is made. A consumed final comparison is preserved in the research repo; |
| future tuning needs a fresh, independently labeled final partition. |
|
|
| ## License and links |
|
|
| Apache-2.0 for the runtime and PaddleOCR-derived models. See LICENSE and NOTICE |
| for PaddlePaddle / RapidAI attribution and the recognizer modification notice. |
| No affiliation with or endorsement by PaddlePaddle or RapidAI is implied. |
|
|
| - [GitHub source and full evaluation reports](https://github.com/MMWilliams/tinyocr) |
| - [PyPI tinyocr-onnx](https://pypi.org/project/tinyocr-onnx/) |
| - [Original detector card](https://huggingface.co/PaddlePaddle/PP-OCRv6_tiny_det_onnx) |
| - [Original recognizer card](https://huggingface.co/PaddlePaddle/PP-OCRv6_tiny_rec_onnx) |
|
|
