<!-- more -->

### 引言

来自 PaddleOCR[官方文档](https://www.paddleocr.ai/latest/version3.x/algorithm/PP-OCRv6/PP-OCRv6.html)：

> PP-OCRv6 是 PP-OCR 最新一代通用文字识别解决方案。PP-OCRv6 基于全新设计的 PPLCNetV4 统一骨干网络，提供 tiny, small, medium 三档模型，分别面向端侧 /IoT、移动端 / 桌面端、服务端场景。PP-OCRv6 在语言覆盖方面实现重大突破，medium/small 档单一模型统一支持简体中文、繁体中文、英文、日文及 46 种拉丁语系语言共 50 种语言（tiny 档支持 49 种，不含日文）。在内部多场景综合评估集上，PP-OCRv6_medium 相比 PP-OCRv5_server 识别精度提升 5.1%、检测精度提升 4.6%，同时 GPU 推理速度提升 2.37×；以仅 34.5M 参数的规模，精度超越 Qwen3-VL-235B, GPT-5.5 等大型视觉语言模型。

官方模型托管地址：https://www.modelscope.cn/collections/PaddlePaddle/PP-OCRv6

### 以下代码运行环境

- OS: macOS Tahoe 26.5.1
- Python: 3.10.14
- PaddlePaddle: 3.1.0
- paddle2onnx: 2.1.0
- paddlex: 3.7.1
- rapidocr: 3.8.4

### 1. 模型跑通

该步骤主要先基于 PaddleX 可以正确使用 PP-OCRv6_medium_det 模型得到正确结果。

该部分主要参考文档：[docs](https://paddlepaddle.github.io/PaddleX/latest/module_usage/tutorials/ocr_modules/text_detection.html)

安装 `paddlex`:

```bash linenums="1"
pip install "paddlex[ocr]==3.7.1"
```

测试 PP-OCRv6_medium_det 模型能否正常识别：

!!! tip

    运行以下代码时，模型会自动下载到 **/Users/用户名/.paddlex/official_models** 下。

测试图：[link](https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.png)

```python linenums="1"
from paddlex import create_model

# medium
model = create_model(model_name="PP-OCRv6_medium_det")

# small
model = create_model(model_name="PP-OCRv6_small_det")

# tiny
model = create_model(model_name="PP-OCRv6_tiny_det")

output = model.predict("images/general_ocr_001.png", batch_size=1)
for res in output:
    res.print()
    res.save_to_img(save_path="./output/")
    res.save_to_json(save_path="./output/res.json")
```

预期结果如下，表明成功运行：

![PP-OCRv6_medium_det_general_ocr_001_res](../images/PP-OCRv6_medium_det_general_ocr_001_res.png)

### 2. 模型转换

PaddlePaddle 官方提供了 ONNX 模型，但是考虑到自己训练的模型，仍然需要转换。因此这一步更多地是验证使用当前工具可以自行转换模型。

该部分主要参考文档：[docs](https://paddlepaddle.github.io/PaddleX/latest/pipeline_deploy/paddle2onnx.html?h=paddle2onnx#22)

=== "转换 PP-OCRv6_medium_det"

    PaddleX 官方集成了 paddle2onnx 的转换代码：

    ```bash linenums="1"
    paddlex --install paddle2onnx
    pip install onnx==1.17.0

    paddlex --paddle2onnx --paddle_model_dir models/official_models/PP-OCRv6_medium_det --onnx_model_dir models/PP-OCRv6_det_medium
    ```

    输出日志如下，表明转换成功：

    ```bash linenums="1"
    Input dir: models/official_models/PP-OCRv6_medium_det
    Output dir: models/PP-OCRv6_det_medium
    Paddle2ONNX conversion starting...
    /Users/xxxx/miniconda3/envs/py310/lib/python3.10/site-packages/paddle/utils/cpp_extension/extension_utils.py:715: UserWarning: No ccache found. Please be aware that recompiling all source files may be required. You can download and install ccache from: https://github.com/ccache/ccache/blob/master/doc/INSTALL.md
    warnings.warn(warning_message)
    [Paddle2ONNX] Start parsing the Paddle model file...
    [Paddle2ONNX] Use opset_version = 11 for ONNX export.
    [Paddle2ONNX] PaddlePaddle model is exported as ONNX format now.
    2026-06-17 22:13:44 [INFO]      Try to perform constant folding on the ONNX model with Polygraphy.
    [W] 'colored' module is not installed, will not use colors when logging. To enable colors, please install the 'colored' module: python3 -m pip install colored
    [I] Folding Constants | Pass 1
    [I]     Total Nodes | Original:  1268, After Folding:   597 |   671 Nodes Folded
    [I] Folding Constants | Pass 2
    [I]     Total Nodes | Original:   597, After Folding:   597 |     0 Nodes Folded
    2026-06-17 22:13:53 [INFO]      ONNX model saved in models/PP-OCRv6_det_medium/inference.onnx.
    Paddle2ONNX conversion succeeded
    Copied models/official_models/PP-OCRv6_medium_det/inference.yml to models/PP-OCRv6_det_medium/inference.yml
    Done
    ```

=== "转换 PP-OCRv6_small_det"

    PaddleX 官方集成了 paddle2onnx 的转换代码：

    ```bash linenums="1"
    paddlex --install paddle2onnx
    pip install onnx==1.17.0

    paddlex --paddle2onnx --paddle_model_dir models/official_models/PP-OCRv6_small_det --onnx_model_dir models/PP-OCRv6_det_small
    ```

    输出日志如下，表明转换成功：

    ```bash linenums="1"
    Input dir: models/official_models/PP-OCRv6_small_det
    Output dir: models/PP-OCRv6_det_small
    Paddle2ONNX conversion starting...
    /Users/xxxx/miniconda3/envs/py310/lib/python3.10/site-packages/paddle/utils/cpp_extension/extension_utils.py:715: UserWarning: No ccache found. Please be awarethat recompiling all source files may be required. You can download and install ccache from: https://github.com/ccache/ccache/blob/master/doc/INSTALL.md
    warnings.warn(warning_message)
    [Paddle2ONNX] Start parsing the Paddle model file...
    [Paddle2ONNX] Use opset_version = 11 for ONNX export.
    [Paddle2ONNX] PaddlePaddle model is exported as ONNX format now.
    2026-06-17 22:20:02 [INFO]      Try to perform constant folding on the ONNX model with Polygraphy.
    [W] 'colored' module is not installed, will not use colors when logging. To enablecolors, please install the 'colored' module: python3 -m pip install colored
    [I] Folding Constants | Pass 1
    [I]     Total Nodes | Original:   928, After Folding:   464 |   464 Nodes Folded
    [I] Folding Constants | Pass 2
    [I]     Total Nodes | Original:   464, After Folding:   464 |     0 Nodes Folded
    2026-06-17 22:20:11 [INFO]      ONNX model saved in models/PP-OCRv6_det_small/inference.onnx.
    Paddle2ONNX conversion succeeded
    Copied models/official_models/PP-OCRv6_small_det/inference.yml to models/PP-OCRv6_det_small/inference.yml
    Done
    ```

=== "转换 PP-OCRv6_tiny_det"

    PaddleX 官方集成了 paddle2onnx 的转换代码：

    ```bash linenums="1"
    paddlex --install paddle2onnx
    pip install onnx==1.17.0

    paddlex --paddle2onnx --paddle_model_dir models/official_models/PP-OCRv6_tiny_det --onnx_model_dir models/PP-OCRv6_det_tiny
    ```

    输出日志如下，表明转换成功：

    ```bash linenums="1"
    Input dir: models/official_models/PP-OCRv6_tiny_det
    Output dir: models/PP-OCRv6_det_tiny
    Paddle2ONNX conversion starting...
    /Users/xxxx/miniconda3/envs/py310/lib/python3.10/site-packages/paddle/utils/cpp_extension/extension_utils.py:715: UserWarning: No ccache found. Please be aware that recompiling all source files may be required. You can download and install ccache from: https://github.com/ccache/ccache/blob/master/doc/INSTALL.md
    warnings.warn(warning_message)
    [Paddle2ONNX] Start parsing the Paddle model file...
    [Paddle2ONNX] Use opset_version = 11 for ONNX export.
    [Paddle2ONNX] PaddlePaddle model is exported as ONNX format now.
    2026-06-17 22:21:03 [INFO]      Try to perform constant folding on the ONNX model with Polygraphy.
    [W] 'colored' module is not installed, will not use colors when logging. To enable colors, please install the 'colored' module: python3 -m pip install colored
    [I] Folding Constants | Pass 1
    [I]     Total Nodes | Original:   928, After Folding:   464 |   464 Nodes Folded
    [I] Folding Constants | Pass 2
    [I]     Total Nodes | Original:   464, After Folding:   464 |     0 Nodes Folded
    2026-06-17 22:21:11 [INFO]      ONNX model saved in models/PP-OCRv6_det_tiny/inference.onnx.
    Paddle2ONNX conversion succeeded
    Copied models/official_models/PP-OCRv6_tiny_det/inference.yml to models/PP-OCRv6_det_tiny/inference.yml
    Done
    ```

### 3. 模型推理验证

我这里主要验证 PP-OCRv6_medium_det 模型，small 和 tiny 版除了参数量区别外，其余都一样，因此不做重复验证。

该部分主要是在 RapidOCR 项目中测试能否直接使用 onnx 模型。要点主要是确定模型前后处理是否兼容。从 PaddleOCR config 文件中比较 [PP-OCRv5 mobile det](https://github.com/PaddlePaddle/PaddleOCR/blob/549d83a88b7c75144120e6ec03de80d3eb9e48a5/configs/det/PP-OCRv5/PP-OCRv5_mobile_det.yml) 和 [PP-OCRv6 medium det](https://github.com/PaddlePaddle/PaddleOCR/blob/ef346e0b402934477489001a4d253a20dbeb72a5/configs/det/PP-OCRv6/PP-OCRv6_medium_det.yml) 文件差异：

![v5_mobile_vs_v6_medium_config](../images/v5_mobile_vs_v6_medium_config.jpg)

从上图中可以看出，除了训练阶段的配置有差异外，推理阶段配置基本一模一样，因此现有 `rapidocr` 前后推理代码可以直接使用。

值得注意的是，后处理阶段的中默认参数变了。这一点我这里先不动，后期我再仔细核验。

```python linenums="1"
from rapidocr import RapidOCR

model_path = "models/onnx/PP-OCRv6_det_medium.onnx"
engine = RapidOCR(params={"Det.model_path": model_path})

img_url = "https://paddle-model-ecology.bj.bcebos.com/paddlex/imgs/demo_image/general_ocr_001.png"
result = engine(img_url, use_det=True, use_rec=False, use_cls=False)
print(result)

result.vis("vis_result.jpg")
```

![v6_medium_det_vis_result](../images/v6_medium_det_vis_result.jpg)

### 4. 模型精度测试

!!! warning

    测试集 [text_det_test_dataset](https://huggingface.co/datasets/SWHL/text_det_test_dataset) 包括卡证类、文档类和自然场景三大类。其中卡证类有 82 张，文档类有 75 张，自然场景类有 55 张。缺少手写体、繁体、日文、古籍文本、拼音、艺术字等数据。因此，该基于该测评集的结果仅供参考。

    欢迎有兴趣的小伙伴，可以和我们一起共建更加全面的测评集。

该部分主要使用 [TextDetMetric](https://github.com/SWHL/TextDetMetric) 和测试集 [text_det_test_dataset](https://huggingface.co/datasets/SWHL/text_det_test_dataset) 来评测。

需要安装的包如下：

```bash linenums="1"
pip install datasets
pip install text_det_metric
```

⚠️注意：以下代码基于 `rapidocr==3.8.4` 版本测试

相关测试步骤请参见 [TextDetMetric](https://github.com/SWHL/TextRecMetric) 的 README，一步一步来就行。

=== "(Exp1) PaddleX 框架 + Paddle 格式模型"

    ```python linenums="1" hl_lines="9"
    import time
    import cv2
    import numpy as np
    from datasets import load_dataset
    from tqdm import tqdm

    from paddlex import create_model

    # PP-OCRv6_medium_det / PP-OCRv6_small_det / PP-OCRv6_tiny_det
    model = create_model(model_name="PP-OCRv6_medium_det")

    dataset = load_dataset("SWHL/text_det_test_dataset")
    test_data = dataset["test"]

    content = []
    for i, one_data in enumerate(tqdm(test_data)):
        img = np.array(one_data.get("image"))
        img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)

        t0 = time.perf_counter()
        ocr_results = next(model.predict(input=img, batch_size=1))
        dt_boxes = ocr_results["dt_polys"].tolist()

        elapse = time.perf_counter() - t0

        gt_boxes = [v["points"] for v in one_data["shapes"]]
        content.append(f"{dt_boxes}\t{gt_boxes}\t{elapse}")

    with open("pred.txt", "w", encoding="utf-8") as f:
        for v in content:
            f.write(f"{v}\n")

    from text_det_metric import TextDetMetric

    metric = TextDetMetric()
    pred_path = "pred.txt"
    metric = metric(pred_path)
    print(metric)
    ```

=== "(Exp2) RapidOCR 框架 + Paddle 格式模型"

    ```python linenums="1" hl_lines="11"
    import cv2
    import numpy as np
    from rapidocr import EngineType, OCRVersion, RapidOCR
    from tqdm import tqdm

    from datasets import load_dataset

    model_dir = "models/official_models/PP-OCRv6_medium_det"
    engine = RapidOCR(
        params={
            "Det.ocr_version": OCRVersion.PPOCRV5,
            "Det.engine_type": EngineType.PADDLE,
            "Det.model_dir": model_dir,
        }
    )

    dataset = load_dataset("SWHL/text_det_test_dataset")
    test_data = dataset["test"]

    content = []
    for i, one_data in enumerate(tqdm(test_data)):
        img = np.array(one_data.get("image"))
        img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)

        ocr_results = engine(img, use_det=True, use_cls=False, use_rec=False)
        dt_boxes = ocr_results.boxes

        dt_boxes = [] if dt_boxes is None else dt_boxes.tolist()
        elapse = ocr_results.elapse

        gt_boxes = [v["points"] for v in one_data["shapes"]]
        content.append(f"{dt_boxes}\t{gt_boxes}\t{elapse}")

    with open("pred.txt", "w", encoding="utf-8") as f:
        for v in content:
            f.write(f"{v}\n")

    from text_det_metric import TextDetMetric

    metric = TextDetMetric()
    pred_path = "pred.txt"
    metric = metric(pred_path)
    print(metric)
    ```

=== "(Exp3) RapidOCR 框架 + ONNX Runtime 格式模型"

    ```python linenums="1" hl_lines="11"
    import cv2
    import numpy as np
    from rapidocr import EngineType, ModelType, OCRVersion, RapidOCR
    from tqdm import tqdm

    from datasets import load_dataset

    model_path = "models/onnx/PP-OCRv6_det_medium.onnx"
    engine = RapidOCR(
        params={
            "Det.ocr_version": OCRVersion.PPOCRV5,
            "Det.engine_type": EngineType.ONNXRUNTIME,
            "Det.model_path": model_path,
        }
    )

    dataset = load_dataset("SWHL/text_det_test_dataset")
    test_data = dataset["test"]

    content = []
    for i, one_data in enumerate(tqdm(test_data)):
        img = np.array(one_data.get("image"))
        img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)

        ocr_results = engine(img, use_det=True, use_cls=False, use_rec=False)
        dt_boxes = ocr_results.boxes

        dt_boxes = [] if dt_boxes is None else dt_boxes.tolist()
        elapse = ocr_results.elapse

        gt_boxes = [v["points"] for v in one_data["shapes"]]
        content.append(f"{dt_boxes}\t{gt_boxes}\t{elapse}")

    with open("pred.txt", "w", encoding="utf-8") as f:
        for v in content:
            f.write(f"{v}\n")

    from text_det_metric import TextDetMetric

    metric = TextDetMetric()
    pred_path = "pred.txt"
    metric = metric(pred_path)
    print(metric)
    ```

指标汇总如下（以下指标均为 CPU 下计算所得）：

|Exp|模型|推理框架|模型格式|Precision↑|Recall↑|H-mean↑|Elapse↓|
|:---:|:---|:---|:---|:---:|:---:|:---:|:---:|
|1|PP-OCRv6_medium_det| PaddleX |PaddlePaddle|0.8437|0.8598|0.8517|1.1844|
|2|PP-OCRv6_small_det| PaddleX |PaddlePaddle|0.8374|0.8338|0.8356|0.2745|
|3|PP-OCRv6_tiny_det| PaddleX |PaddlePaddle|0.7599|0.8125|0.7853|0.1403|
|||||||||
|4|PP-OCRv6_medium_det| RapidOCR| PaddlePaddle|0.8254|0.8598|0.8423|3.286|
|5|PP-OCRv6_small_det| RapidOCR| PaddlePaddle|0.854|0.8445|0.8492|0.743|
|6|PP-OCRv6_tiny_det| RapidOCR| PaddlePaddle|0.8244|0.8285|0.8264|0.393|
|||||||||
|7|PP-OCRv6_medium_det| RapidOCR| ONNX Runtime |0.8251|0.8598|0.8421|0.9491|
|8|PP-OCRv6_small_det| RapidOCR| ONNX Runtime |0.854|0.8445|0.8492|0.2277|
|9|PP-OCRv6_tiny_det| RapidOCR| ONNX Runtime |0.8241|0.8285|0.8263|0.1318|
|||||||||
|10|PP-OCRv5_mobile_det| PaddleX |PaddlePaddle|0.7864|0.8018|0.7940|0.1956|
|11|PP-OCRv5_mobile_det| RapidOCR| PaddlePaddle|0.7861|0.8266|0.8058|0.5328|
|12|PP-OCRv5_mobile_det| RapidOCR| ONNX Runtime|0.7861|0.8266|0.8058|0.1653|
|13|PP-OCRv5_mobile_det| RapidOCR| PyTorch|0.7861|0.8266|0.8058|0.8861|
|14|PP-OCRv4_mobile_det| RapidOCR |ONNX Runtime|0.8301|0.8659|0.8476|-|
|||||||||
|15|PP-OCRv5_server_det| PaddleX |PaddlePaddle|0.8347|0.8583|0.8463|2.1450|
|16|PP-OCRv5_server_det| RapidOCR |PaddlePaddle|||||
|17|PP-OCRv5_server_det| RapidOCR| ONNX Runtime|0.7394|0.8442|0.7883|2.0628|
|18|PP-OCRv4_server_det| RapidOCR |ONNX Runtime|0.7922|0.8128|0.7691|-|
|19|PP-OCRv4_server_det| RapidOCR |PyTorch|0.7394|0.8442|0.7883|5.9122|

从以上结果来看，可以得到以下结论：

1. Exp1-3 和 Exp4-6 相比，差异点在于前后处理以及默认参数，从 H-mean 来看，medium 官方推理为 0.8517，而基于 RapidOCR 框架是 0.8423。small 和 tiny 两个模型，H-mean 对比，基于 RapidOCR 框架推理反而更高。**这个具体原因暂时不明。**
2. Exp1-3 和 Exp4-6 推理速度来看，猜测 RapidOCR 中封装的推理 Paddle 格式模型代码应该有些没对齐的，后续这块也会详细查看。
3. Exp4-6 和 Exp7-9 相比，paddle 格式模型转换为 ONNX 格式后，指标几乎一致，说明 **模型转换前后，误差较小，推理速度也有提升**。
4. 从 H-mean 指标来看，Exp8 PP-OCRv6_small_det 中 H-mean 0.8492 要好于 Exp12 PP-OCRv5_mobile_det 0.8058, Exp14 PP-OCRv4_mobile_det 0.8476。同时，考虑到 PP-OCRv6 ONNX 模型的存储大小：medium 62.12 MB, small 9.93 MB, tiny 1.83 MB。**后续集成版本，我会倾向于将 PP-OCRv6_small_det 作为默认模型。**

### 写在最后

这部分代码会集成到 `rapidocr==3.9.0` 中。相关工作正在进行中，欢迎持续关注。