算力卡上部署OCR文本识别服务与测试

使用modelscope上的图像文本行检测和文本识别模型进行本地部署并转为API服务。

本地部署时把代码中的检测和识别模型路径改为本地模型的路径。

关于模型和代码原理可以参见modelscope上这两个模型相关的页面:

iic/cv_resnet18_ocr-detection-db-line-level_damo

iic/cv_convnextTiny_ocr-recognition-handwritten_damo

部署测试ocr模型的图片:

算力卡信息:

python 复制代码
ixsmi
Timestamp    Wed May 28 17:28:09 2025
+-----------------------------------------------------------------------------+
|  IX-ML: 4.1.3       Driver Version: 4.1.3       CUDA Version: 10.2          |
|-------------------------------+----------------------+----------------------|
| GPU  Name                     | Bus-Id               | Clock-SM  Clock-Mem  |
| Fan  Temp  Perf  Pwr:Usage/Cap|      Memory-Usage    | GPU-Util  Compute M. |
|===============================+======================+======================|
| 0    Iluvatar MR-V50A         | 00000000:11:00.0     | 1000MHz   1600MHz    |
| 15%  45C   P0    19W / 75W    | 12290MiB / 16384MiB  | 0%        Default    |
+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU        PID      Process name                                Usage(MiB) |
|=============================================================================|
|    0    2505472      /usr/local/bin/python3 -c from multipro...  864        |
|    0    2503897      python3 ocr_api.py                          256        |
|    0    1688541      /usr/local/bin/python3 -c from multipro...  10992      |
+-----------------------------------------------------------------------------+

注意:以下ocr模型服务代码与硬件平台无关,只要把依赖软件安装了都能运行,即使cpu也能运行。部署测试过程中可能会报缺软件包的问题,根据提示pip install安装后即可运行。

python 复制代码
from fastapi import FastAPI, File, UploadFile
from fastapi.responses import JSONResponse
from modelscope.pipelines import pipeline
from modelscope.utils.constant import Tasks
import uvicorn
import numpy as np
import cv2
import math
from typing import List
from io import BytesIO

# 初始化OCR模型
ocr_detection = pipeline(Tasks.ocr_detection, model='iic/cv_resnet18_ocr-detection-db-line-level_damo')
ocr_recognition = pipeline(Tasks.ocr_recognition, model='iic/cv_convnextTiny_ocr-recognition-handwritten_damo')

app = FastAPI(title="OCR API")

# 工具函数
def crop_image(img, position):
    def distance(x1, y1, x2, y2):
        return math.sqrt((x1 - x2)**2 + (y1 - y2)**2)
    position = position.tolist()
    for i in range(4):
        for j in range(i + 1, 4):
            if position[i][0] > position[j][0]:
                position[i], position[j] = position[j], position[i]
    if position[0][1] > position[1][1]:
        position[0], position[1] = position[1], position[0]
    if position[2][1] > position[3][1]:
        position[2], position[3] = position[3], position[2]
    x1, y1 = position[0]
    x2, y2 = position[2]
    x3, y3 = position[3]
    x4, y4 = position[1]
    corners = np.array([[x1, y1], [x2, y2], [x4, y4], [x3, y3]], dtype=np.float32)
    width = distance((x1 + x4)/2, (y1 + y4)/2, (x2 + x3)/2, (y2 + y3)/2)
    height = distance((x1 + x2)/2, (y1 + y2)/2, (x4 + x3)/2, (y4 + y3)/2)
    dst_corners = np.array([[0, 0], [width-1, 0], [0, height-1], [width-1, height-1]], dtype=np.float32)
    transform = cv2.getPerspectiveTransform(corners, dst_corners)
    dst = cv2.warpPerspective(img, transform, (int(width), int(height)))
    return dst

def order_point(coor):
    arr = np.array(coor).reshape([4, 2])
    centroid = np.mean(arr, axis=0)
    theta = np.arctan2(arr[:, 1] - centroid[1], arr[:, 0] - centroid[0])
    sort_points = arr[np.argsort(theta)]
    if sort_points[0][0] > centroid[0]:
        sort_points = np.concatenate([sort_points[3:], sort_points[:3]])
    return sort_points.astype('float32')

def sort_boxes(boxes):
    def box_center(box):
        x = np.mean([p[0] for p in box])
        y = np.mean([p[1] for p in box])
        return x, y
    centers = [box_center(box) for box in boxes]
    boxes_with_center = list(zip(boxes, centers))
    boxes_with_center.sort(key=lambda x: (x[1][1], x[1][0]))
    return [b[0] for b in boxes_with_center]

# 主OCR函数
def ocr_from_bytes(image_bytes: bytes) -> str:
    image = cv2.imdecode(np.frombuffer(image_bytes, np.uint8), cv2.IMREAD_COLOR)
    det_result = ocr_detection(image)['polygons']
    boxes = [order_point(box) for box in det_result]
    boxes = sort_boxes(boxes)

    lines: List[str] = []
    for pts in boxes:
        crop = crop_image(image, pts)
        text_result = ocr_recognition(crop)
        text = text_result['text'] if isinstance(text_result['text'], str) else ''.join(text_result['text'])
        lines.append(text)
    return '\n'.join(lines)

# FastAPI 路由
@app.post("/ocr")
async def ocr_api(file: UploadFile = File(...)):
    try:
        image_bytes = await file.read()
        result = ocr_from_bytes(image_bytes)
        return JSONResponse(content={"text": result})
    except Exception as e:
        return JSONResponse(content={"error": str(e)}, status_code=500)

# 启动方式(仅用于本地运行时)
# uvicorn ocr_api:app --reload
if __name__ == "__main__":
    uvicorn.run("ocr_api:app", host="0.0.0.0", port=8005, reload=True)

测试:

python 复制代码
import requests

# === 1. API 地址 === 
url = "http://localhost:8005/ocr"  # 改成你的 API 地址

# === 2. 图片路径 === 
image_path = "ocr_img.jpg"  # 本地图片路径

# === 3. 构造请求 === 
with open(image_path, "rb") as f:
    files = {'file': f}
    response = requests.post(url, files=files)

# === 4. 输出结果 === if response.status_code == 200:
    result = response.json()
    print("识别结果:", result.get("text")) else:
    print(f"请求失败,状态码: {response.status_code}")
    print(response.text)

测试结果:

图片:

上面那个"妈妈说..."

测试返回:

约1秒

相关推荐
pingzhuyan10 小时前
python入门篇12-虚拟环境conda的安装与使用
python·ai·llm·ocr·conda
DogDaoDao1 天前
GitHub开源项目Zerox:AI驱动的OCR革命
人工智能·深度学习·开源·github·ocr·图像识别·zerox
程序视点1 天前
望言OCR视频字幕提取2025终极评测:免费版VS专业版提全方位对比(含免费下载)
ocr·视频字幕提取软件·望言ocr·硬字幕识别工具·开源ocr工具·字幕提取方法·视频转文字软件
开开心心_Every2 天前
多线程语音识别工具
javascript·人工智能·ocr·excel·语音识别·symfony
Hi202402172 天前
支持OCR和AI解释的Web PDF阅读器:解决大文档阅读难题
pdf·flask·llm·ocr·阅读器
kevin 12 天前
如何识别发票特殊版式?OCR大模型如何颠覆传统并保证准确率?
ocr
愿你天黑有灯下雨有伞2 天前
告别复杂配置!Spring Boot优雅集成百度OCR的终极方案
spring boot·百度·ocr
胡耀超3 天前
基于Docker的GPU版本飞桨PaddleOCR部署深度指南(国内镜像)2025年7月底测试好用:从理论到实践的完整技术方案
运维·python·docker·容器·ocr·paddlepaddle·gpu
2301_787725654 天前
通用图片 OCR 到 Word API 数据接口
ocr
EkihzniY5 天前
护照阅读器在上海迪士尼乐园
ocr