2026最新OCR识别源码：文字提取与表格还原的深度实践指南

1# Tesseract 5.5基础识别代码（2026最新版）
2import pytesseract
3from PIL import Image
4
5def extract_text(image_path):
6    config = r'--oem 3 --psm 6 -c tessedit_char_whitelist=0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ'
7    img = Image.open(image_path).convert('L')
8    text = pytesseract.image_to_string(img, config=config, lang='chi_sim+eng')
9    return text.strip()
10

1# PaddleOCR表格还原服务调用示例
2from paddleocr import PaddleOCR, draw_ocr
3import cv2
4
5def extract_table(image_path):
6    ocr = PaddleOCR(use_angle_cls=True, lang="ch", table_engine_type="PP-StructureV2")
7    result = ocr.ocr(image_path, cls=True, table=True)
8    
9    # 绘制表格结构（可视化验证）
10    image = cv2.imread(image_path)
11    boxes = [line[0] for line in result[0]['html'][1:]]
12    for box in boxes:
13        cv2.polylines(image, [np.int32(box)], True, (0, 255, 0), 2)
14    cv2.imwrite('table_structure.jpg', image)
15    
16    return result[0]['html']  # 返回HTML格式表格数据
17

1def preprocess_image(image_path):
2    img = cv2.imread(image_path)
3    
4    # 1. 智能去噪（2026年改进版非局部均值去噪）
5    denoised = cv2.fastNlMeansDenoisingColored(img, None, 10, 10, 7, 21)
6    
7    # 2. 动态对比度增强
8    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
9    enhanced = clahe.apply(cv2.cvtColor(denoised, cv2.COLOR_BGR2GRAY))
10    
11    # 3. 倾斜校正（基于霍夫变换的改进算法）
12    edges = cv2.Canny(enhanced, 50, 150)
13    lines = cv2.HoughLinesP(edges, 1, np.pi/180, 100, minLineLength=100, maxLineGap=10)
14    if lines is not None:
15        angles = [np.arctan2(line[0][1]-line[0][3], line[0][0]-line[0][2])*180/np.pi for line in lines]
16        median_angle = np.median(angles)
17        (h, w) = enhanced.shape[:2]
18        center = (w // 2, h // 2)
19        M = cv2.getRotationMatrix2D(center, median_angle, 1.0)
20        rotated = cv2.warpAffine(enhanced, M, (w, h), flags=cv2.INTER_CUBIC, borderMode=cv2.BORDER_REPLICATE)
21        return rotated
22    return enhanced
23

1前端（React 19） → WebAssembly版Tesseract.js → 
2   ↓（WebSocket）
3服务端（Go 1.22） → 
4   ├─ 简单请求：PaddleOCR轻量模型
5   └─ 复杂请求：PaddleOCR全量模型 + 结构化后处理
6   ↓（Kafka）
7数据仓库（ClickHouse 23.8）
8

1def refine_table_content(html_str):
2    from bs4 import BeautifulSoup
3    soup = BeautifulSoup(html_str, 'html.parser')
4    
5    # 1. 合并跨行单元格（处理<td rowspan="2">）
6    for td in soup.find_all('td'):
7        if 'rowspan' in td.attrs:
8            rowspan = int(td['rowspan'])
9            del td['rowspan']
10            # 实现跨行逻辑（此处省略具体实现）
11    
12    # 2. 数据类型推断
13    for table in soup.find_all('table'):
14        for row in table.find_all('tr'):
15            for i, td in enumerate(row.find_all('td')):
16                content = td.get_text().strip()
17                if content.replace('.','').isdigit():
18                    td['data-type'] = 'number'
19                elif content.lower() in ['true', 'false']:
20                    td['data-type'] = 'boolean'
21                # 其他类型判断...
22    
23    return str(soup)
24

1# 批量识别示例（PaddleOCR）
2def batch_recognize(image_paths):
3    ocr = PaddleOCR()
4    results = []
5    for path in image_paths:
6        results.append(ocr.ocr(path, cls=True))
7    return results
8