版本導致內存泄漏

processing

.gitignore

@@ -121,6 +121,7 @@ celerybeat.pid
 *.sage.py
 
 # Environments
+data
 .env
 .venv
 env/
@@ -129,6 +130,8 @@ ENV/
 env.bak/
 venv.bak/
 
+database
+
 # Spyder project settings
 .spyderproject
 .spyproject

.vscode/settings.json

@@ -0,0 +1,3 @@
+{
+    "editor.inlineSuggest.showToolbar": "onHover"
+}

README.md

@@ -1,3 +1,11 @@
-# ocr
+# OCR
+
+基于深度学习的文字识别提取标记
+- 由于当前没有较优的语言分类识别方案, 使用四倍算力换精度
+- 当前支持 英文 中文 日文 韩文 俄文 的识别
+- 去除纯数字和单字符以及置信度低于80的文字
+- 数据转json存储于mysql web_images 每张图像对应的 text 字段
+- 文字以空格分隔合并为字符串加入 Elasticsearch 索引
+
+勿使用 paddleocr==2.9.1 存在顯存泄漏問題, 應使用 paddleocr==2.7.3
 
-基于深度学习的文字识别提取标记

main.py

@@ -1,55 +1,59 @@
-import io
+import json
-import oss2
+import base64
-import requests
-
-from PIL import Image, ImageFile
-
-ImageFile.LOAD_TRUNCATED_IMAGES = True
-
-# 下载图片(使用OSS下载)
-def download_image(url:str) -> Image:
-    if url.startswith('http://image.gameuiux.cn/') or url.startswith('https://image.gameuiux.cn/'):
-        try:
-            url = url.replace('http://image.gameuiux.cn/', '').replace('https://image.gameuiux.cn/', '')
-            oss2.defaults.connection_pool_size = 100
-            oss_host = 'oss-cn-shanghai-internal.aliyuncs.com'
-            oss_auth = oss2.Auth('LTAI4GH3qP6VA3QpmTYCgXEW', 'r2wz4bJty8iYfGIcFmEqlY1yon2Ruy')
-            return Image.open(io.BytesIO(oss2.Bucket(oss_auth, f'http://{oss_host}', 'gameui-image2').get_object(url).read()))
-        except Exception:
-            return None
-    else:
-        try:
-            response = requests.get(url)
-            return Image.open(io.BytesIO(response.content))
-        except Exception:
-            print('图片下载失败:', url)
-            return None
-
-
 import pymysql
-import pymysql.cursors
+import requests
+import dotenv
 
-conn = pymysql.connect(host='172.21.216.35', user='gameui', password='gameui@2022', database='gameui', cursorclass=pymysql.cursors.DictCursor)
+CONFIG = dotenv.dotenv_values(".env")
-cursor = conn.cursor()
+user = "admin"
-cursor.execute("SELECT id, content FROM web_images LIMIT 10")
+password = "Complexpass#123"
+bas64encoded_creds = base64.b64encode(bytes(f"{user}:{password}", "utf-8")).decode("utf-8")
+headers = {"Content-type": "application/json", "Authorization": f"Basic {bas64encoded_creds}"}
+index = "images"
+zinc_host = "https://zincsearch.gameui.net"
+zinc_url = f"{zinc_host}/api/{index}/_doc"
 
-# 获取查询结果
+# 将数据刷入zinc, 并保持同步更新
-rows = cursor.fetchall()
+# 如果SQL中某一条数据被删除, 那么zinc中也要删除
-for row in rows:
-    print(row)
-    image = download_image(row.content)
 
-# 关闭游标和连接
+def connect_to_mysql():
-cursor.close()
+    return pymysql.connect(host=CONFIG['MYSQL_HOST'], user=CONFIG['MYSQL_USER'], password=CONFIG['MYSQL_PASSWORD'], database=CONFIG['MYSQL_NAME'], cursorclass=pymysql.cursors.SSDictCursor)
-conn.close()
+
+# 查询已存在的数据写入zinc  LIMIT 0,10
+def query_data(conn):
+    with conn.cursor(pymysql.cursors.SSCursor) as cursor:
+        cursor.execute("SELECT id, text FROM web_images WHERE text!='' AND text!='[]'")
+        for id, text in cursor:
+            data = { "_id": str(id), "text": " ".join([item['text'] for item in json.loads(text)]) }
+            res = requests.put(zinc_url, headers=headers, data=json.dumps(data), proxies={'http': '', 'https': ''})
+            print("\033[1;32m{}\033[0m".format(id) if json.loads(res.text)['message'] == 'ok' else id, data['text'])
+
+query_data(connect_to_mysql())
 
 
-'''
+# TODO 数据被删除时, zinc中也要删除
-from cnocr import CnOcr
+# TODO 可以监听SQL日志, 一旦有数据变动, 就更新zinc
+# TODO 为数据之间建立事件关联, 当删除一条图像数据时, 也要删除对应的图像
 
-img_fp = './x.jpg'
+## 查询数据
-ocr = CnOcr(rec_model_name='ch_PP-OCRv3')  # 所有参数都使用默认值
+#query = {
-out = ocr.ocr(img_fp)
+#    "query": {
-
+#        "bool": {
-print(out)
+#            "must": [
-'''
+#                {
+#                    "query_string": {
+#                        "query": "City:是否"
+#                    }
+#                }
+#            ]
+#        }
+#    },
+#    "sort": [
+#        "-@timestamp"
+#    ],
+#    "from": 0,
+#    "size": 100
+#}
+#zinc_url = zinc_host + "/es/" + index + "/_search"
+#res = requests.post(zinc_url, headers=headers, data=json.dumps(query), proxies={'http': '', 'https': ''})
+#print(json.dumps(json.loads(res.text), indent=4, ensure_ascii=False))

pp.py

@@ -0,0 +1,177 @@
+#!/usr/bin/env python3.10
+
+import gc
+import os
+import io
+import oss2
+import time
+import json
+import base64
+import dotenv
+import pymysql
+import requests
+import numpy as np
+import warnings
+import logging
+import paddle
+
+from PIL import Image, ImageFile
+from paddleocr import PaddleOCR
+
+paddle.set_flags({'FLAGS_fraction_of_gpu_memory_to_use': 0.4}) # 限制显存占用为GPU的80%
+logging.disable(logging.DEBUG)    # 关闭DEBUG日志的打印
+logging.disable(logging.WARNING)  # 关闭WARNING日志的打印
+warnings.filterwarnings("ignore")
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+oss2.defaults.connection_pool_size = 100
+
+config = dotenv.dotenv_values(".env")
+user = config['ZINCSEARCH_USER']
+password = config['ZINCSEARCH_PASSWORD']
+zinc_host = config['ZINCSEARCH_HOST']
+index = config['ZINCSEARCH_INDEX']
+bas64encoded_creds = base64.b64encode(bytes(f"{user}:{password}", "utf-8")).decode("utf-8")
+headers = {"Content-type": "application/json", "Authorization": f"Basic {bas64encoded_creds}"}
+zinc_url = f"{zinc_host}/api/{index}/_doc"
+
+
+class MyEncoder(json.JSONEncoder):
+    def default(self, obj):
+        if isinstance(obj, np.float32):
+            return int(obj)
+        if isinstance(obj, np.ndarray):
+            return obj.astype(int).tolist()
+        return super(MyEncoder, self).default(obj)
+
+
+def download_image(url: str, max_size=32767) -> Image.Image:
+    if url.endswith('.gif') or url.endswith('.GIF'):
+        print(f'跳过GIF {url}')
+        return None
+    try:
+        if url.startswith('http://image.gameuiux.cn/') or url.startswith('https://image.gameuiux.cn/'):
+            url = url.replace('http://image.gameuiux.cn/', '').replace('https://image.gameuiux.cn/', '')
+            if os.path.exists(url):
+                print(f'从本地读取图片 {url}')
+                img = Image.open(url)
+            else:
+                print(f'从OSS下载图片 {url}')
+                oss_auth = oss2.Auth(config['OSS_ACCESS_KEY_ID'], config['OSS_ACCESS_KEY_SECRET'])
+                bucket = oss2.Bucket(oss_auth, f'http://{config["OSS_HOST"]}', config['OSS_BUCKET_NAME'])
+                img = Image.open(io.BytesIO(bucket.get_object(url).read()))
+        else:
+            print(f'从网络下载图片 {url}')
+            response = requests.get(url)
+            img = Image.open(io.BytesIO(response.content))
+        if img.mode != 'RGB':
+            img = img.convert('RGB')
+        if max(img.size) > max_size:
+            print(f'跳过尺寸过大的图像 {url}')
+            return None
+        return img
+    except Exception as e:
+        print(f'图片从{url}下载失败，错误信息为：{e}')
+        return None
+
+
+def connect_to_mysql():
+    return pymysql.connect(host=config['MYSQL_HOST'], user=config['MYSQL_USER'], password=config['MYSQL_PASSWORD'], database=config['MYSQL_NAME'], cursorclass=pymysql.cursors.SSDictCursor)
+
+
+
+# 中英日韩俄
+EN = PaddleOCR(use_angle_cls=True, lang="en")
+CH = PaddleOCR(use_angle_cls=True, lang="ch")
+JP = PaddleOCR(use_angle_cls=True, lang="japan")
+KR = PaddleOCR(use_angle_cls=True, lang="korean")
+RU = PaddleOCR(use_angle_cls=True, lang="ru")
+
+# 运行OCR并清理内存
+def process_ocr(model, image):
+    result = model.ocr(image, cls=True)[0] or []
+    paddle.device.cuda.empty_cache()  # 清理缓存
+    gc.collect()  # 强制垃圾回收
+    return result
+
+def process_images(conn, offset=0) -> int:
+    global EN, CH, JP, KR, RU
+    with conn.cursor(pymysql.cursors.SSCursor) as cursor:
+        cursor.execute("SELECT id, content FROM web_images WHERE text='' AND article_category_top_id=22 LIMIT 100 OFFSET %s", (offset,))
+        for id, content in cursor.fetchall():
+            image = download_image(content)
+            if image is None:
+                continue
+            if isinstance(image, Image.Image):
+                image = np.array(image)
+            print(id, content)
+
+            # 執行提取文字
+            #en = EN.ocr(image, cls=True)[0] or []
+            #ch = CH.ocr(image, cls=True)[0] or []
+            #jp = JP.ocr(image, cls=True)[0] or []
+            #kr = KR.ocr(image, cls=True)[0] or []
+            #ru = RU.ocr(image, cls=True)[0] or []
+            # 处理每个模型
+            ru = process_ocr(RU, image)
+            en = process_ocr(EN, image)
+            ch = process_ocr(CH, image)
+            jp = process_ocr(JP, image)
+            kr = process_ocr(KR, image)
+
+            # 排除字符长度小于2的行, 排除纯数字的行, 排除置信度小于 0.8 的行
+            jp = [x for x in jp if len(x[1][0]) > 1 and not x[1][0].isdigit() and x[1][1] > 0.8]
+            kr = [x for x in kr if len(x[1][0]) > 1 and not x[1][0].isdigit() and x[1][1] > 0.8]
+            ch = [x for x in ch if len(x[1][0]) > 1 and not x[1][0].isdigit() and x[1][1] > 0.8]
+            en = [x for x in en if len(x[1][0]) > 1 and not x[1][0].isdigit() and x[1][1] > 0.8]
+            ru = [x for x in ru if len(x[1][0]) > 1 and not x[1][0].isdigit() and x[1][1] > 0.8]
+            print(f'置信度大于 0.8 的行: jp {len(jp)} kr {len(kr)} ch {len(ch)} en {len(en)} ru {len(ru)}')
+
+            # 去除字符串中包含的数字和标点(不作计数)
+            jp_ex = [[x[0], (x[1][0].translate(str.maketrans('', '', '0123456789.,，。！？：；“”‘’\'\"')), x[1][1])] for x in jp]
+            kr_ex = [[x[0], (x[1][0].translate(str.maketrans('', '', '0123456789.,，。！？：；“”‘’\'\"')), x[1][1])] for x in kr]
+            ch_ex = [[x[0], (x[1][0].translate(str.maketrans('', '', '0123456789.,，。！？：；“”‘’\'\"')), x[1][1])] for x in ch]
+            en_ex = [[x[0], (x[1][0].translate(str.maketrans('', '', '0123456789.,，。！？：；“”‘’\'\"')), x[1][1])] for x in en]
+            ru_ex = [[x[0], (x[1][0].translate(str.maketrans('', '', '0123456789.,，。！？：；“”‘’\'\"')), x[1][1])] for x in ru]
+
+            # 计算置信度平均值 x 计算总字数
+            jpx = (np.mean([x[1][1] for x in jp_ex]) if jp_ex else 0) * len(''.join([x[1][0] for x in jp_ex]))
+            krx = (np.mean([x[1][1] for x in kr_ex]) if kr_ex else 0) * len(''.join([x[1][0] for x in kr_ex]))
+            chx = (np.mean([x[1][1] for x in ch_ex]) if ch_ex else 0) * len(''.join([x[1][0] for x in ch_ex]))
+            enx = (np.mean([x[1][1] for x in en_ex]) if en_ex else 0) * len(''.join([x[1][0] for x in en_ex]))
+            rux = (np.mean([x[1][1] for x in ru_ex]) if ru_ex else 0) * len(''.join([x[1][0] for x in ru_ex]))
+
+            # 找出置信度最高的语言, 结构化存储
+            confidences = {'jp': jpx, 'kr': krx, 'ch': chx, 'en': enx, 'ru': rux}
+            max_confidence_language = max(confidences, key=confidences.get)
+            languages = {'en': en, 'ch': ch, 'jp': jp, 'kr': kr, 'ru': ru}
+            data = [{'text': text[0], 'confidence': text[1], 'coordinate': coord} for coord, text in languages[max_confidence_language]]
+            #print("data:", data)
+
+            # 转换为字符串存储到索引库
+            obj = { "_id": str(id), "text": ' '.join([x['text'] for x in data]) }
+            print("转换为字符串存储到索引库:", obj)
+            res = requests.put(zinc_url, headers=headers, data=json.dumps(obj), proxies={'http': '', 'https': ''})
+            print("\033[1;32m{}\033[0m".format(id) if json.loads(res.text)['message'] == 'ok' else obj["id"], obj["text"])
+
+            # 转换为 JSON 存储到数据库
+            with conn.cursor() as c:
+                data = json.dumps(data, ensure_ascii=False, cls=MyEncoder)
+                c.execute("UPDATE web_images SET text = %s WHERE id = %s", (data, id))
+                conn.commit()
+            paddle.device.cuda.empty_cache()  # 清理缓存
+            gc.collect()  # 强制垃圾回收
+    paddle.device.cuda.empty_cache()  # 清理缓存
+    gc.collect()  # 强制垃圾回收
+    return offset+100
+
+def main():
+    conn = connect_to_mysql()
+    offset = 2000
+    while True:
+        print("LOOP:", offset)
+        offset = process_images(conn, offset)
+        time.sleep(0)
+
+
+if __name__ == "__main__":
+    main()

requirements.txt

@@ -1,82 +1,9 @@
-aiohttp==3.8.6
+wheel==0.45.0
-aiosignal==1.3.1
+numpy==1.26.2
-appdirs==1.4.4
+oss2==2.18.3
-async-timeout==4.0.3
+paddleocr==2.7.3
-attrs==23.1.0
+paddlepaddle-gpu=2.6.2
-certifi==2023.7.22
-charset-normalizer==3.3.2
-click==8.1.7
-cnocr==2.2.4.2
-cnstd==1.2.3.5
-coloredlogs==15.0.1
-contourpy==1.2.0
-cycler==0.12.1
-docker-pycreds==0.4.0
-filelock==3.13.1
-flatbuffers==23.5.26
-fonttools==4.44.0
-frozenlist==1.4.0
-fsspec==2023.10.0
-gitdb==4.0.11
-GitPython==3.1.40
-huggingface-hub==0.19.0
-humanfriendly==10.0
-idna==3.4
-Jinja2==3.1.2
-kiwisolver==1.4.5
-lightning-utilities==0.9.0
-MarkupSafe==2.1.3
-matplotlib==3.8.1
-mpmath==1.3.0
-multidict==6.0.4
-networkx==3.2.1
-numpy==1.26.1
-nvidia-cublas-cu12==12.1.3.1
-nvidia-cuda-cupti-cu12==12.1.105
-nvidia-cuda-nvrtc-cu12==12.1.105
-nvidia-cuda-runtime-cu12==12.1.105
-nvidia-cudnn-cu12==8.9.2.26
-nvidia-cufft-cu12==11.0.2.54
-nvidia-curand-cu12==10.3.2.106
-nvidia-cusolver-cu12==11.4.5.107
-nvidia-cusparse-cu12==12.1.0.106
-nvidia-nccl-cu12==2.18.1
-nvidia-nvjitlink-cu12==12.3.52
-nvidia-nvtx-cu12==12.1.105
-onnx==1.15.0
-onnxruntime==1.16.2
-opencv-python==4.8.1.78
-packaging==23.2
-pandas==2.1.3
 Pillow==10.1.0
-Polygon3==3.0.9.1
-protobuf==4.25.0
-psutil==5.9.6
-pyclipper==1.3.0.post5
 PyMySQL==1.1.0
-pyparsing==3.1.1
+python-dotenv==1.0.0
-python-dateutil==2.8.2
+Requests==2.31.0
-pytorch-lightning==2.1.1
-pytz==2023.3.post1
-PyYAML==6.0.1
-requests==2.31.0
-scipy==1.11.3
-seaborn==0.13.0
-sentry-sdk==1.34.0
-setproctitle==1.3.3
-shapely==2.0.2
-six==1.16.0
-smmap==5.0.1
-sympy==1.12
-torch==2.1.0+cpu
-torchaudio==2.1.0
-torchmetrics==1.2.0
-torchvision==0.16.0+cpu
-tqdm==4.66.1
-triton==2.1.0
-typing_extensions==4.8.0
-tzdata==2023.3
-Unidecode==1.3.7
-urllib3==2.0.7
-wandb==0.16.0
-yarl==1.9.2