Bug: 一个1.44K的空白线条形状的16G内存运行PP-OCRv5会内存溢出

[dragons96]

问题描述 / Problem Description

该图片运行在16G Linux服务器上会因为内存溢出程序终止

运行环境 / Runtime Environment

Ubuntu 24.04.2 LTS Linux 6.8.0-57-generic 8C16G
python3.12.10
rapidocr 3.3.1

复现代码 / Reproduction Code

import time

from loguru import logger
from rapidocr import RapidOCR
from rapidocr.utils.process_img import ResizeImgError
from enum import Enum

class OcrVersion(str, Enum):
    pp_ocrv4 = "PP-OCRv4"
    pp_ocrv5 = "PP-OCRv5"


class RapidOCRClient:
    """Rapid OCR client"""

    def __init__(self, version: OcrVersion = None):
        if version is None:
            self.engine = RapidOCR()
        else:
            self.engine = RapidOCR(
                params={
                    'Det.ocr_version': version,
                    'Rec.ocr_version': version,
                }
            )

    def predict(self, image: str) -> list[str]:
        start_time = time.time()
        try:
            result = self.engine(image)
            return list(result.txts) if result.txts else []
        except ResizeImgError:
            logger.info(f'图片像素过少, 无法计算OCR计算, 图片地址: {image}')
            return []
        except Exception as e:
            logger.error(f'图片OCR计算异常, 异常信息: {str(e)}, 图片地址: {image}')
            return []
        finally:
            end_time = time.time()
            logger.info('图片: {} OCR计算耗时: {:.2f} 秒'.format(image, end_time - start_time))


if __name__ == '__main__':
    ans = RapidOCRClient(version=OcrVersion.pp_ocrv5).predict('image.jpg')
    print(ans)

可能解决方案 / Possible solutions

[SWHL]

这个是因为检测模型没有限制最大边。可以尝试更改：

RapidOCR/python/rapidocr/config.yaml

Lines 78 to 79 in a0bb926

	limit_side_len: 736
	limit_type: min

为：

    limit_side_len: 4000
    limit_type: max