paddlex 3.0.0rc1__py3-none-any.whl → 3.0.2__py3-none-any.whl

paddlex/inference/pipelines/ocr/pipeline.py

@@ -22,6 +22,7 @@ from ...common.batch_sampler import ImageBatchSampler
 from ...common.reader import ReadImage
 from ...utils.hpi import HPIConfig
 from ...utils.pp_option import PaddlePredictorOption
+from .._parallel import AutoParallelImageSimpleInferencePipeline
 from ..base import BasePipeline
 from ..components import (
     CropByPolys,
@@ -33,12 +34,9 @@ from ..components import (
 from .result import OCRResult
 
 
-@pipeline_requires_extra("ocr")
-class OCRPipeline(BasePipeline):
+class _OCRPipeline(BasePipeline):
     """OCR Pipeline"""
 
-    entities = "OCR"
-
     def __init__(
         self,
         config: Dict,
@@ -55,9 +53,9 @@ class OCRPipeline(BasePipeline):
             device (str, optional): Device to run the predictions on. Defaults to None.
             pp_option (PaddlePredictorOption, optional): PaddlePredictor options. Defaults to None.
             use_hpip (bool, optional): Whether to use the high-performance
-                inference plugin (HPIP). Defaults to False.
+                inference plugin (HPIP) by default. Defaults to False.
             hpi_config (Optional[Union[Dict[str, Any], HPIConfig]], optional):
-                The high-performance inference configuration dictionary.
+                The default high-performance inference configuration dictionary.
                 Defaults to None.
         """
         super().__init__(
@@ -93,6 +91,7 @@ class OCRPipeline(BasePipeline):
         if self.text_type == "general":
             self.text_det_limit_side_len = text_det_config.get("limit_side_len", 960)
             self.text_det_limit_type = text_det_config.get("limit_type", "max")
+            self.text_det_max_side_limit = text_det_config.get("max_side_limit", 4000)
             self.text_det_thresh = text_det_config.get("thresh", 0.3)
             self.text_det_box_thresh = text_det_config.get("box_thresh", 0.6)
             self.input_shape = text_det_config.get("input_shape", None)
@@ -102,6 +101,7 @@ class OCRPipeline(BasePipeline):
         elif self.text_type == "seal":
             self.text_det_limit_side_len = text_det_config.get("limit_side_len", 736)
             self.text_det_limit_type = text_det_config.get("limit_type", "min")
+            self.text_det_max_side_limit = text_det_config.get("max_side_limit", 4000)
             self.text_det_thresh = text_det_config.get("thresh", 0.2)
             self.text_det_box_thresh = text_det_config.get("box_thresh", 0.6)
             self.text_det_unclip_ratio = text_det_config.get("unclip_ratio", 0.5)
@@ -115,6 +115,7 @@ class OCRPipeline(BasePipeline):
             text_det_config,
             limit_side_len=self.text_det_limit_side_len,
             limit_type=self.text_det_limit_type,
+            max_side_limit=self.text_det_max_side_limit,
             thresh=self.text_det_thresh,
             box_thresh=self.text_det_box_thresh,
             unclip_ratio=self.text_det_unclip_ratio,
@@ -131,7 +132,7 @@ class OCRPipeline(BasePipeline):
             text_rec_config, input_shape=self.input_shape
         )
 
-        self.batch_sampler = ImageBatchSampler(batch_size=1)
+        self.batch_sampler = ImageBatchSampler(batch_size=config.get("batch_size", 1))
         self.img_reader = ReadImage(format="BGR")
 
     def rotate_image(
@@ -234,6 +235,7 @@ class OCRPipeline(BasePipeline):
         self,
         text_det_limit_side_len: Optional[int] = None,
         text_det_limit_type: Optional[str] = None,
+        text_det_max_side_limit: Optional[int] = None,
         text_det_thresh: Optional[float] = None,
         text_det_box_thresh: Optional[float] = None,
         text_det_unclip_ratio: Optional[float] = None,
@@ -246,6 +248,7 @@ class OCRPipeline(BasePipeline):
         Args:
             text_det_limit_side_len (Optional[int]): The maximum side length of the text box.
             text_det_limit_type (Optional[str]): The type of limit to apply to the text box.
+            text_det_max_side_limit (Optional[int]): The maximum side length of the text box.
             text_det_thresh (Optional[float]): The threshold for text detection.
             text_det_box_thresh (Optional[float]): The threshold for the bounding box.
             text_det_unclip_ratio (Optional[float]): The ratio for unclipping the text box.
@@ -257,6 +260,8 @@ class OCRPipeline(BasePipeline):
             text_det_limit_side_len = self.text_det_limit_side_len
         if text_det_limit_type is None:
             text_det_limit_type = self.text_det_limit_type
+        if text_det_max_side_limit is None:
+            text_det_max_side_limit = self.text_det_max_side_limit
         if text_det_thresh is None:
             text_det_thresh = self.text_det_thresh
         if text_det_box_thresh is None:
@@ -267,6 +272,7 @@ class OCRPipeline(BasePipeline):
             limit_side_len=text_det_limit_side_len,
             limit_type=text_det_limit_type,
             thresh=text_det_thresh,
+            max_side_limit=text_det_max_side_limit,
             box_thresh=text_det_box_thresh,
             unclip_ratio=text_det_unclip_ratio,
         )
@@ -279,6 +285,7 @@ class OCRPipeline(BasePipeline):
         use_textline_orientation: Optional[bool] = None,
         text_det_limit_side_len: Optional[int] = None,
         text_det_limit_type: Optional[str] = None,
+        text_det_max_side_limit: Optional[int] = None,
         text_det_thresh: Optional[float] = None,
         text_det_box_thresh: Optional[float] = None,
         text_det_unclip_ratio: Optional[float] = None,
@@ -294,6 +301,7 @@ class OCRPipeline(BasePipeline):
             use_textline_orientation (Optional[bool]): Whether to use textline orientation prediction.
             text_det_limit_side_len (Optional[int]): Maximum side length for text detection.
             text_det_limit_type (Optional[str]): Type of limit to apply for text detection.
+            text_det_max_side_limit (Optional[int]): Maximum side length for text detection.
             text_det_thresh (Optional[float]): Threshold for text detection.
             text_det_box_thresh (Optional[float]): Threshold for text detection boxes.
             text_det_unclip_ratio (Optional[float]): Ratio for unclipping text detection boxes.
@@ -312,6 +320,7 @@ class OCRPipeline(BasePipeline):
         text_det_params = self.get_text_det_params(
             text_det_limit_side_len,
             text_det_limit_type,
+            text_det_max_side_limit,
             text_det_thresh,
             text_det_box_thresh,
             text_det_unclip_ratio,
@@ -320,87 +329,135 @@ class OCRPipeline(BasePipeline):
         if text_rec_score_thresh is None:
             text_rec_score_thresh = self.text_rec_score_thresh
 
-        for img_id, batch_data in enumerate(self.batch_sampler(input)):
-            image_array = self.img_reader(batch_data.instances)[0]
+        for _, batch_data in enumerate(self.batch_sampler(input)):
+            image_arrays = self.img_reader(batch_data.instances)
 
             if model_settings["use_doc_preprocessor"]:
-                doc_preprocessor_res = next(
+                doc_preprocessor_results = list(
                     self.doc_preprocessor_pipeline(
-                        image_array,
+                        image_arrays,
                         use_doc_orientation_classify=use_doc_orientation_classify,
                         use_doc_unwarping=use_doc_unwarping,
                     )
                 )
             else:
-                doc_preprocessor_res = {"output_img": image_array}
+                doc_preprocessor_results = [{"output_img": arr} for arr in image_arrays]
 
-            doc_preprocessor_image = doc_preprocessor_res["output_img"]
+            doc_preprocessor_images = [
+                item["output_img"] for item in doc_preprocessor_results
+            ]
 
-            det_res = next(
-                self.text_det_model(doc_preprocessor_image, **text_det_params)
+            det_results = list(
+                self.text_det_model(doc_preprocessor_images, **text_det_params)
             )
 
-            dt_polys = det_res["dt_polys"]
-            det_res["dt_scores"]
-
-            dt_polys = self._sort_boxes(dt_polys)
-
-            single_img_res = {
-                "input_path": batch_data.input_paths[0],
-                "page_index": batch_data.page_indexes[0],
-                "doc_preprocessor_res": doc_preprocessor_res,
-                "dt_polys": dt_polys,
-                "model_settings": model_settings,
-                "text_det_params": text_det_params,
-                "text_type": self.text_type,
-                "text_rec_score_thresh": text_rec_score_thresh,
-            }
-
-            single_img_res["rec_texts"] = []
-            single_img_res["rec_scores"] = []
-            single_img_res["rec_polys"] = []
-            if len(dt_polys) > 0:
-                all_subs_of_img = list(
-                    self._crop_by_polys(doc_preprocessor_image, dt_polys)
+            dt_polys_list = [item["dt_polys"] for item in det_results]
+
+            dt_polys_list = [self._sort_boxes(item) for item in dt_polys_list]
+
+            results = [
+                {
+                    "input_path": input_path,
+                    "page_index": page_index,
+                    "doc_preprocessor_res": doc_preprocessor_res,
+                    "dt_polys": dt_polys,
+                    "model_settings": model_settings,
+                    "text_det_params": text_det_params,
+                    "text_type": self.text_type,
+                    "text_rec_score_thresh": text_rec_score_thresh,
+                    "rec_texts": [],
+                    "rec_scores": [],
+                    "rec_polys": [],
+                }
+                for input_path, page_index, doc_preprocessor_res, dt_polys in zip(
+                    batch_data.input_paths,
+                    batch_data.page_indexes,
+                    doc_preprocessor_results,
+                    dt_polys_list,
                 )
+            ]
+
+            indices = list(range(len(doc_preprocessor_images)))
+            indices = [idx for idx in indices if len(dt_polys_list[idx]) > 0]
+
+            if indices:
+                all_subs_of_imgs = []
+                chunk_indices = [0]
+                for idx in indices:
+                    all_subs_of_img = list(
+                        self._crop_by_polys(
+                            doc_preprocessor_images[idx], dt_polys_list[idx]
+                        )
+                    )
+                    all_subs_of_imgs.extend(all_subs_of_img)
+                    chunk_indices.append(chunk_indices[-1] + len(all_subs_of_img))
+
                 # use textline orientation model
                 if model_settings["use_textline_orientation"]:
                     angles = [
                         int(textline_angle_info["class_ids"][0])
                         for textline_angle_info in self.textline_orientation_model(
-                            all_subs_of_img
+                            all_subs_of_imgs
                         )
                     ]
-                    all_subs_of_img = self.rotate_image(all_subs_of_img, angles)
+                    all_subs_of_imgs = self.rotate_image(all_subs_of_imgs, angles)
                 else:
-                    angles = [-1] * len(all_subs_of_img)
-                single_img_res["textline_orientation_angles"] = angles
-
-                sub_img_info_list = [
-                    {
-                        "sub_img_id": img_id,
-                        "sub_img_ratio": sub_img.shape[1] / float(sub_img.shape[0]),
-                    }
-                    for img_id, sub_img in enumerate(all_subs_of_img)
-                ]
-                sorted_subs_info = sorted(
-                    sub_img_info_list, key=lambda x: x["sub_img_ratio"]
-                )
-                sorted_subs_of_img = [
-                    all_subs_of_img[x["sub_img_id"]] for x in sorted_subs_info
-                ]
-                for idx, rec_res in enumerate(self.text_rec_model(sorted_subs_of_img)):
-                    sub_img_id = sorted_subs_info[idx]["sub_img_id"]
-                    sub_img_info_list[sub_img_id]["rec_res"] = rec_res
-                for sno in range(len(sub_img_info_list)):
-                    rec_res = sub_img_info_list[sno]["rec_res"]
-                    if rec_res["rec_score"] >= text_rec_score_thresh:
-                        single_img_res["rec_texts"].append(rec_res["rec_text"])
-                        single_img_res["rec_scores"].append(rec_res["rec_score"])
-                        single_img_res["rec_polys"].append(dt_polys[sno])
-            if self.text_type == "general":
-                rec_boxes = convert_points_to_boxes(single_img_res["rec_polys"])
-                single_img_res["rec_boxes"] = rec_boxes
-            else:
-                single_img_res["rec_boxes"] = np.array([])
-            yield OCRResult(single_img_res)
+                    angles = [-1] * len(all_subs_of_imgs)
+                for i, idx in enumerate(indices):
+                    res = results[idx]
+                    res["textline_orientation_angles"] = angles[
+                        chunk_indices[i] : chunk_indices[i + 1]
+                    ]
+
+                # TODO: Process all sub-images in the batch together
+                for i, idx in enumerate(indices):
+                    all_subs_of_img = all_subs_of_imgs[
+                        chunk_indices[i] : chunk_indices[i + 1]
+                    ]
+                    res = results[idx]
+                    dt_polys = dt_polys_list[idx]
+                    sub_img_info_list = [
+                        {
+                            "sub_img_id": img_id,
+                            "sub_img_ratio": sub_img.shape[1] / float(sub_img.shape[0]),
+                        }
+                        for img_id, sub_img in enumerate(all_subs_of_img)
+                    ]
+                    sorted_subs_info = sorted(
+                        sub_img_info_list, key=lambda x: x["sub_img_ratio"]
+                    )
+                    sorted_subs_of_img = [
+                        all_subs_of_img[x["sub_img_id"]] for x in sorted_subs_info
+                    ]
+                    for i, rec_res in enumerate(
+                        self.text_rec_model(sorted_subs_of_img)
+                    ):
+                        sub_img_id = sorted_subs_info[i]["sub_img_id"]
+                        sub_img_info_list[sub_img_id]["rec_res"] = rec_res
+                    for sno in range(len(sub_img_info_list)):
+                        rec_res = sub_img_info_list[sno]["rec_res"]
+                        if rec_res["rec_score"] >= text_rec_score_thresh:
+                            res["rec_texts"].append(rec_res["rec_text"])
+                            res["rec_scores"].append(rec_res["rec_score"])
+                            res["rec_polys"].append(dt_polys[sno])
+
+            for res in results:
+                if self.text_type == "general":
+                    rec_boxes = convert_points_to_boxes(res["rec_polys"])
+                    res["rec_boxes"] = rec_boxes
+                else:
+                    res["rec_boxes"] = np.array([])
+
+                yield OCRResult(res)
+
+
+@pipeline_requires_extra("ocr")
+class OCRPipeline(AutoParallelImageSimpleInferencePipeline):
+    entities = "OCR"
+
+    @property
+    def _pipeline_cls(self):
+        return _OCRPipeline
+
+    def _get_batch_size(self, config):
+        return config.get("batch_size", 1)

paddlex/inference/pipelines/ocr/result.py

@@ -14,14 +14,13 @@
 
 import math
 import random
-from pathlib import Path
 from typing import Dict
 
 import numpy as np
 from PIL import Image, ImageDraw
 
 from ....utils.deps import class_requires_deps, function_requires_deps, is_dep_available
-from ....utils.fonts import SIMFANG_FONT_FILE_PATH, create_font
+from ....utils.fonts import SIMFANG_FONT_FILE_PATH, create_font, create_font_vertical
 from ...common.result import BaseCVResult, JsonMixin
 
 if is_dep_available("opencv-contrib-python"):
@@ -32,15 +31,6 @@ if is_dep_available("opencv-contrib-python"):
 class OCRResult(BaseCVResult):
     """OCR result"""
 
-    def _get_input_fn(self):
-        fn = super()._get_input_fn()
-        if (page_idx := self["page_index"]) is not None:
-            fp = Path(fn)
-            stem, suffix = fp.stem, fp.suffix
-            return f"{stem}_{page_idx}{suffix}"
-        else:
-            return fn
-
     def get_minarea_rect(self, points: np.ndarray) -> np.ndarray:
         """
         Get the minimum area rectangle for the given points using OpenCV.
@@ -106,7 +96,9 @@ class OCRResult(BaseCVResult):
                     height = int(0.5 * (max(box[:, 1]) - min(box[:, 1])))
                     box[:2, 1] = np.mean(box[:, 1])
                     box[2:, 1] = np.mean(box[:, 1]) + min(20, height)
-                draw_left.polygon(box, fill=color)
+                box_pts = [(int(x), int(y)) for x, y in box.tolist()]
+                draw_left.polygon(box_pts, fill=color)
+
                 img_right_text = draw_box_txt_fine(
                     (w, h), box, txt, SIMFANG_FONT_FILE_PATH
                 )
@@ -214,19 +206,20 @@ def draw_box_txt_fine(
         np.ndarray: An image with the text drawn in the specified box.
     """
     box_height = int(
-        math.sqrt((box[0][0] - box[3][0]) ** 2 + (box[0][1] - box[3][1]) ** 2)
+        math.sqrt(float(box[0][0] - box[3][0]) ** 2 + float(box[0][1] - box[3][1]) ** 2)
     )
     box_width = int(
-        math.sqrt((box[0][0] - box[1][0]) ** 2 + (box[0][1] - box[1][1]) ** 2)
+        math.sqrt(float(box[0][0] - box[1][0]) ** 2 + float(box[0][1] - box[1][1]) ** 2)
     )
 
     if box_height > 2 * box_width and box_height > 30:
-        img_text = Image.new("RGB", (box_height, box_width), (255, 255, 255))
+        img_text = Image.new("RGB", (box_width, box_height), (255, 255, 255))
         draw_text = ImageDraw.Draw(img_text)
         if txt:
-            font = create_font(txt, (box_height, box_width), font_path)
-            draw_text.text([0, 0], txt, fill=(0, 0, 0), font=font)
-        img_text = img_text.transpose(Image.ROTATE_270)
+            font = create_font_vertical(txt, (box_width, box_height), font_path)
+            draw_vertical_text(
+                draw_text, (0, 0), txt, font, fill=(0, 0, 0), line_spacing=2
+            )
     else:
         img_text = Image.new("RGB", (box_width, box_height), (255, 255, 255))
         draw_text = ImageDraw.Draw(img_text)
@@ -250,3 +243,13 @@ def draw_box_txt_fine(
         borderValue=(255, 255, 255),
     )
     return img_right_text
+
+
+@function_requires_deps("opencv-contrib-python")
+def draw_vertical_text(draw, position, text, font, fill=(0, 0, 0), line_spacing=2):
+    x, y = position
+    for char in text:
+        draw.text((x, y), char, font=font, fill=fill)
+        bbox = font.getbbox(char)
+        char_height = bbox[3] - bbox[1]
+        y += char_height + line_spacing

paddlex/inference/pipelines/open_vocabulary_detection/pipeline.py

@@ -45,9 +45,9 @@ class OpenVocabularyDetectionPipeline(BasePipeline):
             device (str): The device to run the prediction on. Default is None.
             pp_option (PaddlePredictorOption): Options for PaddlePaddle predictor. Default is None.
             use_hpip (bool, optional): Whether to use the high-performance
-                inference plugin (HPIP). Defaults to False.
+                inference plugin (HPIP) by default. Defaults to False.
             hpi_config (Optional[Union[Dict[str, Any], HPIConfig]], optional):
-                The high-performance inference configuration dictionary.
+                The default high-performance inference configuration dictionary.
                 Defaults to None.
         """
         super().__init__(

paddlex/inference/pipelines/open_vocabulary_segmentation/pipeline.py

@@ -47,9 +47,9 @@ class OpenVocabularySegmentationPipeline(BasePipeline):
             device (str): The device to run the prediction on. Default is None.
             pp_option (PaddlePredictorOption): Options for PaddlePaddle predictor. Default is None.
             use_hpip (bool, optional): Whether to use the high-performance
-                inference plugin (HPIP). Defaults to False.
+                inference plugin (HPIP) by default. Defaults to False.
             hpi_config (Optional[Union[Dict[str, Any], HPIConfig]], optional):
-                The high-performance inference configuration dictionary.
+                The default high-performance inference configuration dictionary.
                 Defaults to None.
         """
         super().__init__(

paddlex/inference/pipelines/pp_chatocr/pipeline_base.py

@@ -37,9 +37,9 @@ class PP_ChatOCR_Pipeline(BasePipeline):
             device (str, optional): Device to run the predictions on. Defaults to None.
             pp_option (PaddlePredictorOption, optional): PaddlePredictor options. Defaults to None.
             use_hpip (bool, optional): Whether to use the high-performance
-                inference plugin (HPIP). Defaults to False.
+                inference plugin (HPIP) by default. Defaults to False.
             hpi_config (Optional[Union[Dict[str, Any], HPIConfig]], optional):
-                The high-performance inference configuration dictionary.
+                The default high-performance inference configuration dictionary.
                 Defaults to None.
         """
 

paddlex/inference/pipelines/pp_chatocr/pipeline_v3.py

@@ -54,9 +54,9 @@ class PP_ChatOCRv3_Pipeline(PP_ChatOCR_Pipeline):
             device (str, optional): Device to run the predictions on. Defaults to None.
             pp_option (PaddlePredictorOption, optional): PaddlePredictor options. Defaults to None.
             use_hpip (bool, optional): Whether to use the high-performance
-                inference plugin (HPIP). Defaults to False.
+                inference plugin (HPIP) by default. Defaults to False.
             hpi_config (Optional[Union[Dict[str, Any], HPIConfig]], optional):
-                The high-performance inference configuration dictionary.
+                The default high-performance inference configuration dictionary.
                 Defaults to None.
             initial_predictor (bool, optional): Whether to initialize the predictor. Defaults to True.
         """
@@ -206,7 +206,6 @@ class PP_ChatOCRv3_Pipeline(PP_ChatOCR_Pipeline):
         input: Union[str, List[str], np.ndarray, List[np.ndarray]],
         use_doc_orientation_classify: Optional[bool] = None,
         use_doc_unwarping: Optional[bool] = None,
-        use_general_ocr: Optional[bool] = None,
         use_seal_recognition: Optional[bool] = None,
         use_table_recognition: Optional[bool] = None,
         layout_threshold: Optional[Union[float, dict]] = None,
@@ -237,7 +236,6 @@ class PP_ChatOCRv3_Pipeline(PP_ChatOCR_Pipeline):
                                                                         numpy array of an image, or list of numpy arrays.
             use_doc_orientation_classify (bool): Flag to use document orientation classification.
             use_doc_unwarping (bool): Flag to use document unwarping.
-            use_general_ocr (bool): Flag to use general OCR.
             use_seal_recognition (bool): Flag to use seal recognition.
             use_table_recognition (bool): Flag to use table recognition.
             layout_threshold (Optional[float]): The threshold value to filter out low-confidence predictions. Default is None.
@@ -280,7 +278,6 @@ class PP_ChatOCRv3_Pipeline(PP_ChatOCR_Pipeline):
             input,
             use_doc_orientation_classify=use_doc_orientation_classify,
             use_doc_unwarping=use_doc_unwarping,
-            use_general_ocr=use_general_ocr,
             use_seal_recognition=use_seal_recognition,
             use_table_recognition=use_table_recognition,
             layout_threshold=layout_threshold,

paddlex/inference/pipelines/pp_chatocr/pipeline_v4.py

@@ -62,9 +62,9 @@ class PP_ChatOCRv4_Pipeline(PP_ChatOCR_Pipeline):
             device (str, optional): Device to run the predictions on. Defaults to None.
             pp_option (PaddlePredictorOption, optional): PaddlePredictor options. Defaults to None.
             use_hpip (bool, optional): Whether to use the high-performance
-                inference plugin (HPIP). Defaults to False.
+                inference plugin (HPIP) by default. Defaults to False.
             hpi_config (Optional[Union[Dict[str, Any], HPIConfig]], optional):
-                The high-performance inference configuration dictionary.
+                The default high-performance inference configuration dictionary.
                 Defaults to None.
             initial_predictor (bool, optional): Whether to initialize the predictor. Defaults to True.
         """
@@ -249,7 +249,7 @@ class PP_ChatOCRv4_Pipeline(PP_ChatOCR_Pipeline):
         input: Union[str, List[str], np.ndarray, List[np.ndarray]],
         use_doc_orientation_classify: Optional[bool] = None,
         use_doc_unwarping: Optional[bool] = None,
-        use_general_ocr: Optional[bool] = None,
+        use_textline_orientation: Optional[bool] = None,
         use_seal_recognition: Optional[bool] = None,
         use_table_recognition: Optional[bool] = None,
         layout_threshold: Optional[Union[float, dict]] = None,
@@ -280,7 +280,7 @@ class PP_ChatOCRv4_Pipeline(PP_ChatOCR_Pipeline):
                                                                         numpy array of an image, or list of numpy arrays.
             use_doc_orientation_classify (bool): Flag to use document orientation classification.
             use_doc_unwarping (bool): Flag to use document unwarping.
-            use_general_ocr (bool): Flag to use general OCR.
+            use_textline_orientation (Optional[bool]): Whether to use textline orientation prediction.
             use_seal_recognition (bool): Flag to use seal recognition.
             use_table_recognition (bool): Flag to use table recognition.
             layout_threshold (Optional[float]): The threshold value to filter out low-confidence predictions. Default is None.
@@ -322,7 +322,7 @@ class PP_ChatOCRv4_Pipeline(PP_ChatOCR_Pipeline):
             input,
             use_doc_orientation_classify=use_doc_orientation_classify,
             use_doc_unwarping=use_doc_unwarping,
-            use_general_ocr=use_general_ocr,
+            use_textline_orientation=use_textline_orientation,
             use_seal_recognition=use_seal_recognition,
             use_table_recognition=use_table_recognition,
             layout_threshold=layout_threshold,
@@ -638,7 +638,7 @@ class PP_ChatOCRv4_Pipeline(PP_ChatOCR_Pipeline):
 
         for image_array in self.img_reader([input]):
 
-            image_string = cv2.imencode(".jpg", image_array)[1].tostring()
+            image_string = cv2.imencode(".jpg", image_array)[1].tobytes()
             image_base64 = base64.b64encode(image_string).decode("utf-8")
             result = {}
             for key in key_list:

paddlex/inference/pipelines/rotated_object_detection/pipeline.py

@@ -20,15 +20,13 @@ from ....utils.deps import pipeline_requires_extra
 from ...models.object_detection.result import DetResult
 from ...utils.hpi import HPIConfig
 from ...utils.pp_option import PaddlePredictorOption
+from .._parallel import AutoParallelImageSimpleInferencePipeline
 from ..base import BasePipeline
 
 
-@pipeline_requires_extra("cv")
-class RotatedObjectDetectionPipeline(BasePipeline):
+class _RotatedObjectDetectionPipeline(BasePipeline):
     """Rotated Object Detection Pipeline"""
 
-    entities = "rotated_object_detection"
-
     def __init__(
         self,
         config: Dict,
@@ -45,9 +43,9 @@ class RotatedObjectDetectionPipeline(BasePipeline):
             device (str): The device to run the prediction on. Default is None.
             pp_option (PaddlePredictorOption): Options for PaddlePaddle predictor. Default is None.
             use_hpip (bool, optional): Whether to use the high-performance
-                inference plugin (HPIP). Defaults to False.
+                inference plugin (HPIP) by default. Defaults to False.
             hpi_config (Optional[Union[Dict[str, Any], HPIConfig]], optional):
-                The high-performance inference configuration dictionary.
+                The default high-performance inference configuration dictionary.
                 Defaults to None.
         """
         super().__init__(
@@ -83,3 +81,15 @@ class RotatedObjectDetectionPipeline(BasePipeline):
             DetResult: The predicted rotated object detection results.
         """
         yield from self.rotated_object_detection_model(input, threshold=threshold)
+
+
+@pipeline_requires_extra("cv")
+class RotatedObjectDetectionPipeline(AutoParallelImageSimpleInferencePipeline):
+    entities = "rotated_object_detection"
+
+    @property
+    def _pipeline_cls(self):
+        return _RotatedObjectDetectionPipeline
+
+    def _get_batch_size(self, config):
+        return config["SubModules"]["RotatedObjectDetection"].get("batch_size", 1)