| --- |
| license: apache-2.0 |
| library_name: onnx |
| tags: |
| - ocr |
| - text-detection |
| - paddleocr |
| - onnx |
| base_model: PaddlePaddle/PaddleOCR |
| pipeline_tag: object-detection |
| --- |
| |
| # PaddleOCR PP-OCRv4 — Text Region Detection (ONNX) |
|
|
| ONNX export of PaddlePaddle's [PP-OCRv4](https://github.com/PaddlePaddle/PaddleOCR) text-region detector. Despite the `ch_` prefix in the filename, the detector itself is largely language-agnostic — script-specific weights live in the recognizer, not the detector. |
|
|
| Re-hosted under Heliosoph for distribution stability — PaddleOCR's GitHub is the authoritative source. |
|
|
| Credit: Baidu / PaddlePaddle OCR team. |
|
|
| ## What this repo contains |
|
|
| ``` |
| ch_PP-OCRv4_det.onnx # ~5 MB |
| ``` |
|
|
| The detection stage of a two-stage OCR pipeline. Pair with a recognizer (TrOCR, PP-OCRv4 recognizer, or use Florence-2's `<OCR_WITH_REGION>` mode as an end-to-end alternative). |
|
|
| ## Input/output |
|
|
| | | Spec | |
| |---|---| |
| | Input name | `x` | |
| | Input shape | dynamic — typically `[1, 3, H, W]` where H, W are multiples of 32 | |
| | Input dtype | float32 | |
| | Input color order | **RGB** | |
| | Preprocessing | Resize to nearest multiple of 32 (with max dim cap, e.g. 960), normalize by `mean=[0.485, 0.456, 0.406]`, `std=[0.229, 0.224, 0.225]` | |
| | Output | Probability map `[1, 1, H, W]` | |
| | Postprocessing | Threshold at ~0.3, find connected components, fit minimum-area rectangle or quadrilateral per component | |
|
|
| The output is a per-pixel "is this part of a text region?" probability. The postprocessing turns the heatmap into polygons. |
|
|
| ## How to use |
|
|
| ```python |
| import onnxruntime as ort |
| import numpy as np |
| from PIL import Image |
| |
| sess = ort.InferenceSession("ch_PP-OCRv4_det.onnx") |
| |
| # Resize maintaining aspect ratio, snap to multiples of 32 |
| img = Image.open("document.jpg").convert("RGB") |
| img_resized = resize_to_multiple_of_32(img, max_side=960) # custom helper |
| |
| arr = (np.asarray(img_resized, dtype=np.float32) / 255.0 - [0.485, 0.456, 0.406]) / [0.229, 0.224, 0.225] |
| arr = arr.transpose(2, 0, 1)[None, ...] # 1x3xHxW |
| |
| prob_map = sess.run(None, {"x": arr})[0][0, 0] # HxW probability map |
| |
| # Threshold and extract polygons |
| binary = (prob_map > 0.3).astype(np.uint8) |
| polygons = extract_polygons_from_binary_mask(binary) # custom — see PaddleOCR DBNet postprocess |
| ``` |
|
|
| Reference postprocessing: [PaddleOCR's DBNet decoder](https://github.com/PaddlePaddle/PaddleOCR/blob/main/ppocr/postprocess/db_postprocess.py). |
|
|
| ## When to use this vs Florence-2 OCR |
|
|
| - **PaddleOCR detector + recognizer**: faster, tinier, language-specific recognizer choice, classical two-stage pipeline |
| - **Florence-2 `<OCR_WITH_REGION>`**: single model, broader language coverage, larger compute footprint |
| - **TrOCR**: best for printed Latin text; not language-agnostic |
|
|
| If you're building a document-OCR pipeline today, PP-OCRv4 detection + TrOCR recognition is hard to beat for English. For mixed languages, Florence-2. |
|
|
| ## License |
|
|
| **Apache-2.0** — same as upstream. `LICENSE` file included. |
|
|
