关键信息提取中文本识别的置信度

[eunij-peanut]

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import numpy as np

import os
import sys

__dir__ = os.path.dirname(os.path.abspath(__file__))
sys.path.append(__dir__)
sys.path.insert(0, os.path.abspath(os.path.join(__dir__, "..")))

os.environ["FLAGS_allocator_strategy"] = "auto_growth"
import cv2
import json
import paddle

from ppocr.data import create_operators, transform
from ppocr.modeling.architectures import build_model
from ppocr.postprocess import build_post_process
from ppocr.utils.save_load import load_model
from ppocr.utils.visual import draw_ser_results
from ppocr.utils.utility import get_image_file_list, load_vqa_bio_label_maps
import tools.program as program


def to_tensor(data):
    import numbers
    from collections import defaultdict

    data_dict = defaultdict(list)
    to_tensor_idxs = []

    for idx, v in enumerate(data):
        if isinstance(v, (np.ndarray, paddle.Tensor, numbers.Number)):
            if idx not in to_tensor_idxs:
                to_tensor_idxs.append(idx)
        data_dict[idx].append(v)
    for idx in to_tensor_idxs:
        data_dict[idx] = paddle.to_tensor(data_dict[idx])
    return list(data_dict.values())


class SerPredictor(object):
    def __init__(self, config):
        global_config = config["Global"]
        self.algorithm = config["Architecture"]["algorithm"]

        
        self.post_process_class = build_post_process(
            config["PostProcess"], global_config
        )

        
        self.model = build_model(config["Architecture"])

        load_model(config, self.model, model_type=config["Architecture"]["model_type"])

        from paddleocr import PaddleOCR

        self.ocr_engine = PaddleOCR(
            use_angle_cls=False,
            show_log=False,
            rec_model_dir=global_config.get("kie_rec_model_dir", None),
            det_model_dir=global_config.get("kie_det_model_dir", None),
            use_gpu=global_config["use_gpu"],
        )

        
        transforms = []
        for op in config["Eval"]["dataset"]["transforms"]:
            op_name = list(op)[0]
            if "Label" in op_name:
                op[op_name]["ocr_engine"] = self.ocr_engine
            elif op_name == "KeepKeys":
                op[op_name]["keep_keys"] = [
                    "input_ids",
                    "bbox",
                    "attention_mask",
                    "token_type_ids",
                    "image",
                    "labels",
                    "segment_offset_id",
                    "ocr_info",
                    "entities",
                ]

            transforms.append(op)
        if config["Global"].get("infer_mode", None) is None:
            global_config["infer_mode"] = True
        self.ops = create_operators(
            config["Eval"]["dataset"]["transforms"], global_config
        )
        self.model.eval()

    def __call__(self, data):
        with open(data["img_path"], "rb") as f:
            img = f.read()
        data["image"] = img
        batch = transform(data, self.ops)
        batch = to_tensor(batch)
        preds = self.model(batch)

        post_result = self.post_process_class(
            preds, segment_offset_ids=batch[6], ocr_infos=batch[7]
        )
        return post_result, batch


if __name__ == "__main__":
    config, device, logger, vdl_writer = program.preprocess()
    os.makedirs(config["Global"]["save_res_path"], exist_ok=True)

    ser_engine = SerPredictor(config)

    if config["Global"].get("infer_mode", None) is False:
        data_dir = config["Eval"]["dataset"]["data_dir"]
        with open(config["Global"]["infer_img"], "rb") as f:
            infer_imgs = f.readlines()
    else:
        infer_imgs = get_image_file_list(config["Global"]["infer_img"])

    with open(
        os.path.join(config["Global"]["save_res_path"], "infer_results.txt"),
        "w",
        encoding="utf-8",
    ) as fout:
        for idx, info in enumerate(infer_imgs):
            if config["Global"].get("infer_mode", None) is False:
                data_line = info.decode("utf-8")
                substr = data_line.strip("\n").split("\t")
                img_path = os.path.join(data_dir, substr[0])
                data = {"img_path": img_path, "label": substr[1]}
            else:
                img_path = info
                data = {"img_path": img_path}

            save_img_path = os.path.join(
                config["Global"]["save_res_path"],
                os.path.splitext(os.path.basename(img_path))[0] + "_ser.jpg",
            )

            result, _ = ser_engine(data)
            result = result[0]
            fout.write(
                img_path
                + "\t"
                + json.dumps(
                    {
                        "ocr_info": result,
                    },
                    ensure_ascii=False,
                )
                + "\n"
            )
            img_res = draw_ser_results(img_path, result)
            cv2.imwrite(save_img_path, img_res)

            logger.info(
                "process: [{}/{}], save result to {}".format(
                    idx, len(infer_imgs), save_img_path
                )
            )

这是ser的推理代码，请问我该如何修改，才能在文本识别过程，得到每个文本的置信度？

[GreatV]

要在文本识别过程中获取每个文本的置信度，你需要检查 PaddleOCR 的检测和识别结果，并提取置信度信息。PaddleOCR 的 ocr_engine.ocr() 方法返回的结果通常包含文本框坐标、识别的文本以及对应的置信度。

修改代码以获取文本识别置信度

在 SerPredictor.__call__ 方法中，当前代码已经调用了 PaddleOCR 进行 OCR 识别，但未提取文本置信度。你可以修改 post_result 返回的内容，使其包含 score 信息。

修改 SerPredictor.__call__ 方法

def __call__(self, data):
    with open(data["img_path"], "rb") as f:
        img = f.read()
    data["image"] = img
    batch = transform(data, self.ops)
    batch = to_tensor(batch)
    preds = self.model(batch)

    post_result = self.post_process_class(
        preds, segment_offset_ids=batch[6], ocr_infos=batch[7]
    )

    # 提取 OCR 结果并添加置信度信息
    for item in post_result[0]:  # 遍历检测到的文本区域
        if "score" in item:  # 置信度信息
            item["confidence"] = item["score"]  # 统一命名为 confidence

    return post_result, batch

该修改确保在 post_result 结果中包含 confidence 值，你可以在后续处理中使用该信息。

---

修改 infer_kie_token_ser.py 保存识别结果

在当前代码的 __main__ 函数中，fout.write() 仅保存 OCR 结果的文本信息，你可以修改它以包含文本的置信度：

fout.write(
    img_path
    + "\t"
    + json.dumps(
        {
            "ocr_info": [
                {
                    "text": entry["text"],
                    "confidence": entry["confidence"],  # 添加置信度字段
                    "bbox": entry["bbox"],  # 如果需要位置信息，也可以添加
                }
                for entry in result
            ]
        },
        ensure_ascii=False,
    )
    + "\n"
)

这将确保 JSON 文件中每个文本条目标注其置信度。

---

如何使用置信度过滤无关文本

如果你希望 只保留高置信度的文本，可以在处理 post_result 时加上一个置信度过滤：

threshold = 0.8  # 设定阈值
filtered_result = [
    item for item in post_result[0] if item.get("confidence", 0) >= threshold
]
return filtered_result, batch

这样，只有置信度大于 0.8 的文本才会被保留。

---

参考相关讨论

你的问题与 GitHub Discussions 14182 中的问题相似，该讨论提到了如何筛选出特定类别的信息并忽略无关文本。你可以尝试结合置信度过滤和类别筛选的方式，实现更加精准的关键字段提取。

希望这些修改可以帮助你成功提取文本的置信度并过滤无关内容！ 🚀

---

Response generated by 🤖 feifei-bot | chatgpt-4o-latest

⚠️ This is an AI-generated response. Please verify any critical information.