# RapidOcr


本示例通过RapidOcr模型说明如何使用ONNXRuntime Python API进行图像文本识别模型的推理，包括如何预处理、推理并获取推理结果。



## 模型简介

本示例使用了ch_PP-OCRv3_det + ch_ppocr_mobile_v2.0_cls + ch_PP-OCRv3_rec三个模型，onnx文件在Resource/Models/文件夹下，模型结构可以通过netron (https://netron.app/) 查看，并通过netron查询各个模型的输入输出。
																													
## 预处理

在将数据输入到模型之前，需要对图像做如下预处理操作：
这段代码的目的是在进行字符识别之前，对图像进行预处理，包括读取图像、调整大小、填充、缩放等操作。

本示例代码采用了OpenCV实现了预处理操作：

### TextDetector的预处理

```python
 pre_process_list = {
            "DetResizeForTest": {
                "limit_side_len": config.get("limit_side_len", 736),
                "limit_type": config.get("limit_type", "min"),
            },
            "NormalizeImage": {
                "std": [0.229, 0.224, 0.225],
                "mean": [0.485, 0.456, 0.406],
                "scale": "1./255.",
                "order": "hwc",
            },
            "ToCHWImage": None,
            "KeepKeys": {"keep_keys": ["image", "shape"]},
        }
        self.preprocess_op = create_operators(pre_process_list)

        post_process = {
            "thresh": config.get("thresh", 0.3),
            "box_thresh": config.get("box_thresh", 0.5),
            "max_candidates": config.get("max_candidates", 1000),
            "unclip_ratio": config.get("unclip_ratio", 1.6),
            "use_dilation": config.get("use_dilation", True),
            "score_mode": config.get("score_mode", "fast"),
        }
```
### TextClassifier的预处理

```python
 def resize_norm_img(self, img):
        img_c, img_h, img_w = self.cls_image_shape
        h, w = img.shape[:2]
        ratio = w / float(h)
        if math.ceil(img_h * ratio) > img_w:
            resized_w = img_w
        else:
            resized_w = int(math.ceil(img_h * ratio))

        resized_image = cv2.resize(img, (resized_w, img_h))
        resized_image = resized_image.astype("float32")
        if img_c == 1:
            resized_image = resized_image / 255
            resized_image = resized_image[np.newaxis, :]
        else:
            resized_image = resized_image.transpose((2, 0, 1)) / 255

        resized_image -= 0.5
        resized_image /= 0.5
        padding_im = np.zeros((img_c, img_h, img_w), dtype=np.float32)
        padding_im[:, :, :resized_w] = resized_image
        return padding_im
```
### TextRecognizer的预处理

```python
        def resize_norm_img(self, img, max_wh_ratio):
        img_channel, img_height, img_width = self.rec_image_shape
        assert img_channel == img.shape[2]

        img_width = int(img_height * max_wh_ratio)

        h, w = img.shape[:2]
        ratio = w / float(h)
        if math.ceil(img_height * ratio) > img_width:
            resized_w = img_width
        else:
            resized_w = int(math.ceil(img_height * ratio))

        resized_image = cv2.resize(img, (resized_w, img_height))
        resized_image = resized_image.astype("float32")
        resized_image = resized_image.transpose((2, 0, 1)) / 255
        resized_image -= 0.5
        resized_image /= 0.5

        padding_im = np.zeros((img_channel, img_height, img_width), dtype=np.float32)
        padding_im[:, :, 0:resized_w] = resized_image
        return padding_im
```

## 推理

### 推理分为三部分：

#### 第一部分：

TextDetector使用ch_ppocr_v3_det_infer.onnx模型，这是一个预训练的文本检测模型，用于文本检测任务。它可以检测图像中的文本区域，并返回文本框的位置和边界框信息。

```python

class TextDetector:
    
    ...
    
        post_process = {
            "thresh": config.get("thresh", 0.3),
            "box_thresh": config.get("box_thresh", 0.5),
            "max_candidates": config.get("max_candidates", 1000),
            "unclip_ratio": config.get("unclip_ratio", 1.6),
            "use_dilation": config.get("use_dilation", True),
            "score_mode": config.get("score_mode", "fast"),
        }
        self.postprocess_op = DBPostProcess(**post_process)

        self.infer = OrtInferSession(config)
    
    ...

```

#### 第二部分：

TextClassifier使用ch_ppocr_v2_cls_infer.onnx模型：这是一个预训练的分类器模型，用于文本分类任务。它可以用于判断文本属于哪个类别或类别的概率。

```python
class TextClassifier:

    ...

    def __init__(self, config):
        self.cls_image_shape = config["cls_image_shape"]
        self.cls_batch_num = config["cls_batch_num"]
        self.cls_thresh = config["cls_thresh"]
        self.postprocess_op = ClsPostProcess(config["label_list"])
        self.infer = OrtInferSession(config)

    ...
           
```

#### 第三部分：
TextDetector使用ch_ppocr_v3_rec_infer.onnx：这是一个预训练的文本识别模型，用于文本识别任务。它可以接收一个文本框的图像区域作为输入，并返回该区域中文本的识别

```python
class TextDetector:

    ...
        def __init__(self, config):
        pre_process_list = {
            "DetResizeForTest": {
                "limit_side_len": config.get("limit_side_len", 736),
                "limit_type": config.get("limit_type", "min"),
            },
            "NormalizeImage": {
                "std": [0.229, 0.224, 0.225],
                "mean": [0.485, 0.456, 0.406],
                "scale": "1./255.",
                "order": "hwc",
            },
            "ToCHWImage": None,
            "KeepKeys": {"keep_keys": ["image", "shape"]},
        }
        self.preprocess_op = create_operators(pre_process_list)

        post_process = {
            "thresh": config.get("thresh", 0.3),
            "box_thresh": config.get("box_thresh", 0.5),
            "max_candidates": config.get("max_candidates", 1000),
            "unclip_ratio": config.get("unclip_ratio", 1.6),
            "use_dilation": config.get("use_dilation", True),
            "score_mode": config.get("score_mode", "fast"),
        }
        self.postprocess_op = DBPostProcess(**post_process)
        self.infer = OrtInferSession(config)

    ...

```