main.py

# see huggingface/project/flask_auto_selection.py

# C:\Users\echod\.paddlex\official_models\PP-OCRv5_server_det
# c:\Users\echod\.conda\envs\ppv5\lib\site-packages\paddle\utils\cpp_extension\extension_utils.py 看模型加载的代码在哪
# PP-OCRv5_server_det PP-OCRv5_server_det.yaml  搜这两个
#   paddlex/configs/modules/text_detection/PP-OCRv5_server_det.yaml,sha256=_cS2Eaqb1IJdN0jXPqtc8wsC-gHY0BdS3oOzZfVINCI,1100
#   C:\Users\echod\.conda\envs\ppv5\Lib\site-packages\paddlex\inference\models\text_detection\predictor.py 实际建模好像是这里 
#   Model files already exist.  搜这个
#   C:\Users\echod\.conda\envs\ppv5\Lib\site-packages\paddlex\inference\utils\official_models.py  self._save_dir  'C:/Users/echod/.paddlex/official_models' 
#     要改的是这个目录路径
#   _save_dir = Path(CACHE_DIR) / "official_models"  改 CACHE_DIR 为 相对路径就可以了吧

"""
import os,sys
from pathlib import Path

# 获取python.exe所在目录
python_dir = Path(sys.executable).parent
os.chdir(python_dir)

abs_path = Path(".paddlex").resolve()

DEFAULT_CACHE_DIR = osp.abspath(osp.join(os.path.expanduser("~"), ".paddlex"))
CACHE_DIR = os.environ.get("PADDLE_PDX_CACHE_HOME", DEFAULT_CACHE_DIR)
CACHE_DIR = abs_path
"""


"""

conda create -n ppv5 python==3.10 pip \
  && conda activate ppv5 \
  && python -m pip install paddlepaddle-gpu==3.1.1 -i https://www.paddlepaddle.org.cn/packages/stable/cu118 \
  && pip install paddleocr 
	# python -m pip install paddlepaddle==3.1.1 -i https://www.paddlepaddle.org.cn/packages/stable/cpu/
		# cpu 就这样

# pip install numpy==2.2.4 pillow==11.1.0 protobuf==6.30.2 flask==3.1.2 opencv-python==4.12.0.88 paddlepaddle==3.1.1 paddleocr==3.2.0 --proxy=http://127.0.0.1:7897
    # -i https://mirrors.aliyun.com/pypi/simple/

"""

is_debug = False

dic_cache = {}

from flask import Flask, request, jsonify
import threading
import platform

app = Flask(__name__)

import json
import decimal
import datetime
import base64
import numpy as np
import cv2

from collections import OrderedDict

class DecimalEncoder(json.JSONEncoder):
    def default(self, o):
        if isinstance(o, decimal.Decimal):
            return float(o)
        elif isinstance(o, datetime.datetime):
            return str(o)
        super(DecimalEncoder, self).default(o)

def save_json(filename, dics):
    with open(filename, 'w', encoding='utf-8') as fp:
        json.dump(dics, fp, indent=4, cls=DecimalEncoder, ensure_ascii=False)
        fp.close()

def load_json(filename):
    with open(filename, encoding='utf-8') as fp:
        js = json.load(fp)
        fp.close()
        return js

def base64_to_mat(base64_str):
    """
    将 Base64 字符串转换为 OpenCV Mat 对象（NumPy 数组）
    
    参数:
        base64_str (str): Base64 编码的图片字符串（不可以含前缀如 "data:image/jpeg;base64,"）
    
    返回:
        Mat: OpenCV 图像对象（NumPy 数组），格式为 BGR
    """
    # 处理可能存在的 Base64 前缀（如 "data:image/jpeg;base64,"）
    # if ',' in base64_str:
    #     base64_data = base64_str.split(',')[1]  # 提取纯 Base64 数据部分
    # else:
    # base64_data = base64_str
    
    # 解码 Base64 字符串为二进制字节流
    image_bytes = base64.b64decode(base64_str)
    
    # 将字节流转换为 NumPy 数组（数据类型 uint8）
    nparr = np.frombuffer(image_bytes, np.uint8)
    
    # 使用 OpenCV 解码为 Mat 对象（BGR 格式）
    mat = cv2.imdecode(nparr, cv2.IMREAD_COLOR_BGR)  # cv2.IMREAD_COLOR 保留色彩通道
    
    return mat


from paddleocr import PaddleOCR

ocr = PaddleOCR(
    use_doc_orientation_classify=False,
    use_doc_unwarping=False,
    use_textline_orientation=False)

def ppresult_tojson(img, result):
    global is_debug

    jn = OrderedDict()
    prism_wordsInfo = []
    jn["prism_wordsInfo"] = prism_wordsInfo
    jn["height"] = img.shape[0]
    jn["width"] = img.shape[1]

    for res in result:
        output_img = res['doc_preprocessor_res']['output_img']  # 这是预处理后的图片，坐标可能是这张图的坐标，而且还原不回去
        # img = output_img 
        jsn = res.json['res']
        text_word = jsn['text_word']
        text_word_boxes = jsn['text_word_boxes']
        rec_texts = jsn['rec_texts']
        rec_boxes = jsn['rec_boxes']

        for idx_line, (words, boxs) in enumerate(zip(text_word, text_word_boxes)):
            text_line = rec_texts[idx_line]
            text_box = rec_boxes[idx_line]

            j = OrderedDict()
            prism_wordsInfo.append( j )

            lu = OrderedDict(x=text_box[0], y=text_box[1])
            ru = OrderedDict(x=text_box[2], y=text_box[1])
            rd = OrderedDict(x=text_box[2], y=text_box[3])
            ld = OrderedDict(x=text_box[0], y=text_box[3])

            j["word"] = text_line
            j["pos"] = [ lu, ru, rd, ld ]


            charInfo = []
            j['charInfo'] = charInfo
            j['angle'] = -1
            j["x"] = lu["x"]
            j["y"] = lu["y"]
            j["width"] = ( max(ru["x"], rd["x"])) - ( min(lu["x"], ld["x"]) )
            j["height"] = ( max(ld["y"], rd["y"])) - ( min(lu["y"], ru["y"]) )


            img = cv2.rectangle(img, (lu['x'], lu['y']), (rd['x'], rd['y']), (255, 0, 0), 2)
            if platform.system() == "Windows":
                if is_debug:
                    cv2.imshow('orig', img)
                    cv2.waitKey(0)
                pass
            for idx_word, (word, box) in enumerate(zip(words, boxs)):

                if (len(word) == 1):
                    info = OrderedDict()
                    charInfo.append( info )
                    info["word"] = word
                    info["x"] = box[0]
                    info["y"] = box[1]
                    info["w"] = box[2] - box[0]
                    info["h"] = box[3] - box[1]
                elif (len(word) > 1):
                    for w in word:
                        info = OrderedDict()
                        charInfo.append( info )
                        info["word"] = w
                        info["x"] = box[0]
                        info["y"] = box[1]
                        info["w"] = box[2] - box[0]
                        info["h"] = box[3] - box[1]

                # print(word)
                img = cv2.rectangle(img, (box[0], box[1]), (box[2], box[3]), (0, 255, 0), 2)  # 矩形的左上角, 矩形的右下角
                if platform.system() == "Windows":
                    if is_debug:
                        cv2.imshow('orgin', img)
                        cv2.waitKey(0)
                    pass

        # save_json('out.json', jn)
        break  # 只处理第一张图的结果
    
    return jn

# 限制同时处理的请求数量为1
ocr_semaphore = threading.Semaphore(1)
@app.route('/ppocr', methods=['post'])
def autoselection():
    # request.json 只能够接受方法为POST、Body为raw，header 内容为 application/json类型的数据
    # print(request.json, type(request.json))

    # 使用 request.form 来接受 x-www-form-urlencoded 格式的数据
    # print(request.form, type(request.form))
    
    # form_data = request.form.to_dict()
    # if "img" not in form_data:
    #     return jsonify([])
    
    # base64_str = form_data["img"]

    # 非阻塞方式获取信号量
    if not ocr_semaphore.acquire(blocking=False):
        return jsonify({"warning": "wait pre task done."})
    
    try:

        base64_str = request.json['img']
        
        img = base64_to_mat(base64_str)

        result = ocr.predict(
            input = img, 
            return_word_box=True
            )
            
        jn = ppresult_tojson(img.copy(), result)

        return jsonify(jn)
    
    except Exception as e:
        return jsonify({"error": str(e)})
    finally:
        ocr_semaphore.release()

if __name__ == '__main__':

    if is_debug:
        pth_img = "data/0025.jpg" # "data/第一单元.jpg"

        imgData = np.fromfile(pth_img, dtype=np.uint8)
        img = cv2.imdecode(imgData, cv2.IMREAD_COLOR_BGR)

        # cv2.imshow('orgin', img)
        # cv2.waitKey(0)

        result = ocr.predict(
            input = img, # pth_img, # "data/无标点符号.jpg", 
            return_word_box=True
            )
            
        jn =ppresult_tojson(img.copy(), result)
        save_json('out.json', jn)

            # res.print()       
            # res.save_to_img("output")
            # res.save_to_json("output")
    else:
        app.run(host="0.0.0.0", port=8889, debug=True)