main_v6.py

# see huggingface_echodict/typst-app-clone/tools/extract_pdf_images.py

# see https://huggingface.co/PaddlePaddle/PP-OCRv6_medium_det_safetensors
# see https://huggingface.co/PaddlePaddle/PP-OCRv6_medium_rec_safetensors

# see https://github.com/PaddlePaddle/PaddleOCR/issues/16137  矫正后的图像我们也返回了，坐标是对应于矫正后图像的。另外图像矫正模块是不可逆的，没法把坐标映射回去。想要原图的结果可以关掉图像矫正模块


"""

.venv/Scripts/python post.py
    .venv/bin/python post.py
    测试接口

curl -LsSf https://astral.sh/uv/install.sh | sh
    # linux
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"
    # Windows.

apt-get update && apt-get install -y libgl1-mesa-glx libglib2.0-0 libgl1 libglib2.0-0t64
    # linux

uv venv --python 3.10 --seed --clear

$env:UV_DEFAULT_INDEX="https://pypi.tuna.tsinghua.edu.cn/simple"
    export UV_DEFAULT_INDEX="https://pypi.tuna.tsinghua.edu.cn/simple"
    加速


CPU 版本：

    .venv/Scripts/pip install "paddleocr[all]"
        .venv/bin/pip install "paddleocr[all]"

    .venv/Scripts/pip install transformers --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.tuna.tsinghua.edu.cn/simple
        .venv/bin/pip install transformers --index-url https://download.pytorch.org/whl/cpu --extra-index-url https://pypi.tuna.tsinghua.edu.cn/simple

    .venv/Scripts/pip install paddlepaddle -i https://www.paddlepaddle.org.cn/packages/stable/cpu/
        .venv/bin/pip install paddlepaddle -i https://www.paddlepaddle.org.cn/packages/stable/cpu/

    .venv/Scripts/pip install flask

GPU 版本：

    .venv/Scripts/pip install "paddleocr[all]"
        .venv/bin/pip install "paddleocr[all]"

    .venv/Scripts/pip install paddlepaddle-gpu -i https://www.paddlepaddle.org.cn/packages/stable/cu130/
        .venv/bin/pip install paddlepaddle-gpu==3.3.1 -i https://www.paddlepaddle.org.cn/packages/stable/cu130/
    
    .venv/Scripts/pip install transformers --index-url https://download.pytorch.org/whl/cu130 --extra-index-url https://pypi.tuna.tsinghua.edu.cn/simple
        .venv/bin/pip install transformers --index-url https://download.pytorch.org/whl/cu130 --extra-index-url https://pypi.tuna.tsinghua.edu.cn/simple

    .venv/Scripts/pip install flask
        .venv/bin/pip install flask

"""


is_debug = False
is_debug_api = False

dic_cache = {}

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

app = Flask(__name__)

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

from collections import OrderedDict

def _parse_csv_env(name: str, default: str):
    raw = os.environ.get(name, default)
    parts = [p.strip() for p in raw.split(",")]
    return [p for p in parts if p]

_cors_allowed_origins = set(_parse_csv_env("PPV5_CORS_ALLOWED_ORIGINS", "https://typst-app-clone.pages.dev"))

def _add_cors_headers(resp):
    resp.headers["Access-Control-Allow-Origin"] = "*"
    resp.headers["Access-Control-Allow-Methods"] = "POST, OPTIONS"
    resp.headers["Access-Control-Allow-Headers"] = "Content-Type"
    resp.headers["Access-Control-Allow-Private-Network"] = "true"
    resp.headers["Access-Control-Max-Age"] = "86400"
    return resp

@app.before_request
def _handle_options_preflight():
    if request.method != "OPTIONS":
        return None
    resp = app.make_response(("", 204))
    return _add_cors_headers(resp)

@app.after_request
def _after_request(resp):
    return _add_cors_headers(resp)

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_UNCHANGED)

    if len(mat.shape) != 3:  # 转彩图
        mat = cv2.cvtColor(mat, cv2.COLOR_GRAY2BGR)
    
    return mat


def ppresult_tojson(img, result, reverse=False): # reverse=True 逆序结果行 古籍可能需要逆序
    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，坐标是这张图的坐标，而且还原不回去。关掉图像矫正后坐标就和原图坐标一致了
        # use_doc_unwarping=False。
        # 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']

        if reverse:
            text_word = text_word[::-1]
            text_word_boxes = text_word_boxes[::-1]
            rec_texts = rec_texts[::-1]
            rec_boxes = rec_boxes[::-1]

        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_api:
                    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_api:
                        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

from paddleocr import PaddleOCR
import paddle
print(f"Paddle版本: {paddle.__version__}")
print(f"GPU可用: {paddle.is_compiled_with_cuda()}")
print(f"GPU数量: {paddle.device.cuda.device_count()}")
ocr = PaddleOCR(
    text_detection_model_dir="./PPv6/PP-OCRv6_medium_det_safetensors",
    text_recognition_model_dir="./PPv6/PP-OCRv6_medium_rec_safetensors",
    lang='chinese_cht',                  # 繁体字典
    return_word_box=True,                # 返回每个字符的坐标
    use_doc_orientation_classify=True,   # 整页方向（横/倒）
    use_doc_unwarping=False,             # 关闭弯曲矫正，单字坐标它才准。否则坐标是矫正后图像的坐标
    use_textline_orientation=True,       # 文本行方向分类，竖排靠它
    text_det_thresh=0.1,                 # 默认 0.3 对古籍太高，漏淡墨
    text_det_box_thresh=0.1,             # 同上
    text_rec_score_thresh=0.3,           # 过滤低置信，古籍可放低
)
"""
ch, chinese_cht, en, japan, af, az, bs, ca, cs, cy, da, de, es, et, eu, fi, fr, ga, gl, hr, hu, id, is, it, ku, la, lb, lt, lv, mi, ms, mt, nl, no, oc, pl, pt, qu, rm, ro, rs_latin, sk, sl, sq, sv, sw, tl, tr, uz, vi, french, german
"""

# 限制同时处理的请求数量为1
ocr_semaphore = threading.Semaphore(1)
@app.route('/ppocrv6', methods=['POST', 'OPTIONS'])
@app.route('/ocr', methods=['POST', 'OPTIONS'])
def ppv6():
    # 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"]

    if request.method == "OPTIONS":
        return _add_cors_headers(app.make_response(("", 204)))

    # 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']
        reverse = request.json['reverse'] or False
        
        img = base64_to_mat(base64_str)

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

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

if __name__ == '__main__':

    if is_debug:
        import cv2

        result = ocr.predict("./data/SWX0005_00000_00001.webp")
        for res in result:
            res.print()
            res.save_to_img("output")
            res.save_to_json("output")

            # 单独保存预处理各阶段的图片（均为 BGR 格式）
            pre = res['doc_preprocessor_res']  # pre['output_img'] draw_box.py 画字符框大体上准，但还不太准，坐标对应的就是 output_img , 而不是原图
            cv2.imwrite("output/SWX0005_00000_00001_input_img.png",     pre['input_img'])
            cv2.imwrite("output/SWX0005_00000_00001_rot_img.png",       pre['rot_img'])
            cv2.imwrite("output/SWX0005_00000_00001_output_img.png",    pre['output_img'])  # draw_box.py 
        pass
    else:
        use_https = os.environ.get("PPV5_USE_HTTPS", "").lower() in ("1", "true", "yes", "on")
        host = os.environ.get("PPV5_HOST", "127.0.0.1" if use_https else "0.0.0.0")
        port = int(os.environ.get("PPV5_PORT", "9346"))

        if use_https:
            cert_file = os.environ.get("PPV5_TLS_CERT", os.path.join(".", "certs", "localhost.pem"))
            key_file = os.environ.get("PPV5_TLS_KEY", os.path.join(".", "certs", "localhost-key.pem"))
            if not (os.path.exists(cert_file) and os.path.exists(key_file)):
                raise FileNotFoundError(
                    "TLS 证书不存在。请先用 mkcert 生成并安装根证书，然后生成 localhost 证书：\n"
                    "  mkcert -install\n"
                    "  mkcert -key-file certs\\localhost-key.pem -cert-file certs\\localhost.pem localhost 127.0.0.1 ::1\n"
                    f"当前期望证书路径：{cert_file}\n"
                    f"当前期望私钥路径：{key_file}\n"
                    "也可用环境变量 PPV5_TLS_CERT / PPV5_TLS_KEY 指定实际路径。"
                )
            app.run(host=host, port=port, debug=True, ssl_context=(cert_file, key_file))
        else:
            app.run(host=host, port=port, debug=True)