布局分析技能

概述

本技能利用 surya（一个先进的文档理解系统）进行文档布局分析。能够检测文本块、表格、图形、标题，并确定复杂文档的阅读顺序。

使用方法

提供文档图像或PDF
指定需要检测的布局元素
我将分析结构并返回检测到的区域

示例提示：

"分析此文档页面的布局"
"检测此图像中的所有表格和文本块"
"确定此PDF页面的阅读顺序"
"查找此文档中的标题和段落"

领域知识

surya 基础

from surya.detection import DetectionPredictor
from surya.layout import LayoutPredictor
from surya.reading_order import ReadingOrderPredictor
from PIL import Image

# 加载图像
image = Image.open("document.png")

# 检测布局元素
layout_predictor = LayoutPredictor()
layout_result = layout_predictor([image])

布局元素类型

元素	描述
Text	常规段落文本
Title	文档/章节标题
Section-header	节标题
List-item	项目符号/编号项
Table	表格数据
Figure	图像/图表
Caption	图形/表格标题
Footnote	脚注
Formula	数学公式
Page-header	页眉
Page-footer	页脚

文本检测

from surya.detection import DetectionPredictor
from PIL import Image

# 初始化检测器
detector = DetectionPredictor()

# 加载图像
image = Image.open("document.png")

# 检测文本区域
results = detector([image])

# 访问结果
for page_result in results:
    for bbox in page_result.bboxes:
        print(f"文本区域: {bbox.bbox}")
        print(f"置信度: {bbox.confidence}")

布局分析

from surya.layout import LayoutPredictor
from PIL import Image

# 初始化布局预测器
layout_predictor = LayoutPredictor()

# 分析布局
image = Image.open("document.png")
layout_results = layout_predictor([image])

# 处理结果
for page_result in layout_results:
    for element in page_result.bboxes:
        print(f"类型: {element.label}")
        print(f"边界框: {element.bbox}")
        print(f"置信度: {element.confidence}")

阅读顺序检测

from surya.reading_order import ReadingOrderPredictor
from surya.layout import LayoutPredictor
from PIL import Image

# 先获取布局
layout_predictor = LayoutPredictor()
image = Image.open("document.png")
layout_results = layout_predictor([image])

# 确定阅读顺序
reading_order_predictor = ReadingOrderPredictor()
order_results = reading_order_predictor([image], layout_results)

# 访问有序元素
for page_result in order_results:
    for i, element in enumerate(page_result.ordered_bboxes):
        print(f"{i+1}. {element.label}: {element.bbox}")

带布局的OCR

from surya.ocr import OCRPredictor
from surya.layout import LayoutPredictor
from PIL import Image

# 初始化预测器
ocr_predictor = OCRPredictor()
layout_predictor = LayoutPredictor()

# 加载图像
image = Image.open("document.png")

# 获取布局
layout_results = layout_predictor([image])

# 运行OCR
ocr_results = ocr_predictor([image])

# 合并结果
for layout, ocr in zip(layout_results, ocr_results):
    for layout_elem in layout.bboxes:
        print(f"元素: {layout_elem.label}")
        
        # 查找此布局元素内的OCR文本
        for text_line in ocr.text_lines:
            if boxes_overlap(layout_elem.bbox, text_line.bbox):
                print(f"  文本: {text_line.text}")

处理PDF

from surya.layout import LayoutPredictor
from pdf2image import convert_from_path

def analyze_pdf_layout(pdf_path):
    """分析PDF中所有页面的布局。"""
    
    # 将PDF转换为图像
    images = convert_from_path(pdf_path)
    
    # 初始化预测器
    layout_predictor = LayoutPredictor()
    
    # 分析所有页面
    results = layout_predictor(images)
    
    document_structure = []
    
    for page_num, page_result in enumerate(results):
        page_elements = []
        
        for element in page_result.bboxes:
            page_elements.append({
                'type': element.label,
                'bbox': element.bbox,
                'confidence': element.confidence
            })
        
        document_structure.append({
            'page': page_num + 1,
            'elements': page_elements
        })
    
    return document_structure

structure = analyze_pdf_layout("document.pdf")

可视化

from surya.layout import LayoutPredictor
from PIL import Image, ImageDraw, ImageFont

def visualize_layout(image_path, output_path):
    """可视化检测到的布局元素。"""
    
    image = Image.open(image_path)
    layout_predictor = LayoutPredictor()
    results = layout_predictor([image])
    
    # 创建绘图上下文
    draw = ImageDraw.Draw(image)
    
    # 元素类型颜色映射
    colors = {
        'Text': 'blue',
        'Title': 'red',
        'Table': 'green',
        'Figure': 'purple',
        'Section-header': 'orange',
        'List-item': 'cyan',
    }
    
    for element in results[0].bboxes:
        bbox = element.bbox
        color = colors.get(element.label, 'gray')
        
        # 绘制矩形
        draw.rectangle(bbox, outline=color, width=2)
        
        # 添加标签
        draw.text((bbox[0], bbox[1] - 15), 
                  f"{element.label} ({element.confidence:.2f})",
                  fill=color)
    
    image.save(output_path)
    return output_path

最佳实践

使用高质量图像：建议150 DPI以上以获得最佳效果
必要时预处理：对旋转文档进行去倾斜
验证结果：检查置信度分数
处理多页：逐页处理
结合OCR：获取检测区域内的文本

常见模式

文档结构提取

def extract_document_structure(image_path):
    """提取层次化文档结构。"""
    
    from surya.layout import LayoutPredictor
    from surya.reading_order import ReadingOrderPredictor
    
    image = Image.open(image_path)
    
    # 获取布局
    layout_predictor = LayoutPredictor()
    layout_results = layout_predictor([image])
    
    # 获取阅读顺序
    order_predictor = ReadingOrderPredictor()
    order_results = order_predictor([image], layout_results)
    
    structure = {
        'title': None,
        'sections': [],
        'tables': [],
        'figures': []
    }
    
    current_section = None
    
    for element in order_results[0].ordered_bboxes:
        if element.label == 'Title':
            structure['title'] = element
        elif element.label == 'Section-header':
            current_section = {'header': element, 'content': []}
            structure['sections'].append(current_section)
        elif element.label == 'Table':
            structure['tables'].append(element)
        elif element.label == 'Figure':
            structure['figures'].append(element)
        elif current_section and element.label in ['Text', 'List-item']:
            current_section['content'].append(element)
    
    return structure

表格区域提取

def extract_table_regions(image_path):
    """从文档中提取表格区域。"""
    
    from surya.layout import LayoutPredictor
    
    image = Image.open(image_path)
    layout_predictor = LayoutPredictor()
    results = layout_predictor([image])
    
    tables = []
    
    for element in results[0].bboxes:
        if element.label == 'Table':
            bbox = element.bbox
            
            # 裁剪表格区域
            table_image = image.crop(bbox)
            
            tables.append({
                'bbox': bbox,
                'image': table_image,
                'confidence': element.confidence
            })
    
    return tables

示例

示例1：学术论文分析

from surya.layout import LayoutPredictor
from surya.reading_order import ReadingOrderPredictor
from pdf2image import convert_from_path

def analyze_academic_paper(pdf_path):
    """分析学术论文的结构。"""
    
    images = convert_from_path(pdf_path)
    
    layout_predictor = LayoutPredictor()
    order_predictor = ReadingOrderPredictor()
    
    paper_structure = {
        'pages': [],
        'element_counts': {
            'Title': 0,
            'Section-header': 0,
            'Text': 0,
            'Table': 0,
            'Figure': 0,
            'Formula': 0,
            'Footnote': 0
        }
    }
    
    layout_results = layout_predictor(images)
    order_results = order_predictor(images, layout_results)
    
    for page_num, (layout, order) in enumerate(zip(layout_results, order_results)):
        page_structure = {
            'page': page_num + 1,
            'elements': []
        }
        
        for element in order.ordered_bboxes:
            page_structure['elements'].append({
                'type': element.label,
                'bbox': element.bbox,
                'order': element.position
            })
            
            # 统计元素类型
            if element.label in paper_structure['element_counts']:
                paper_structure['element_counts'][element.label] += 1
        
        paper_structure['pages'].append(page_structure)
    
    return paper_structure

paper = analyze_academic_paper('research_paper.pdf')
print(f"表格总数: {paper['element_counts']['Table']}")
print(f"图形总数: {paper['element_counts']['Figure']}")

示例2：表单字段检测

from surya.layout import LayoutPredictor
from PIL import Image

def detect_form_fields(image_path):
    """检测表单字段和标签。"""
    
    image = Image.open(image_path)
    
    layout_predictor = LayoutPredictor()
    results = layout_predictor([image])
    
    form_fields = []
    
    for element in results[0].bboxes:
        # 查找可能是标签的文本元素
        if element.label == 'Text':
            # 检查附近是否有框/线（可能的输入字段）
            form_fields.append({
                'type': 'potential_label',
                'bbox': element.bbox,
                'confidence': element.confidence
            })
    
    return form_fields

fields = detect_form_fields('form.png')
print(f"找到 {len(fields)} 个潜在表单元素")

示例3：多栏文章

from surya.layout import LayoutPredictor
from surya.reading_order import ReadingOrderPredictor
from PIL import Image

def process_multicolumn_article(image_path):
    """处理多栏文章布局。"""
    
    image = Image.open(image_path)
    
    layout_predictor = LayoutPredictor()
    order_predictor = ReadingOrderPredictor()
    
    layout_results = layout_predictor([image])
    order_results = order_predictor([image], layout_results)
    
    # 按栏分组元素
    image_width = image.width
    column_threshold = image_width / 2
    
    columns = {
        'left': [],
        'right': [],
        'full_width': []
    }
    
    for element in order_results[0].ordered_bboxes:
        bbox = element.bbox
        element_center = (bbox[0] + bbox[2]) / 2
        element_width = bbox[2] - bbox[0]
        
        # 确定栏
        if element_width > column_threshold * 1.5:
            columns['full_width'].append(element)
        elif element_center < column_threshold:
            columns['left'].append(element)
        else:
            columns['right'].append(element)
    
    return {
        'layout': 'multi-column',
        'columns': columns,
        'reading_order': order_results[0].ordered_bboxes
    }

article = process_multicolumn_article('newspaper_page.png')
print(f"左栏: {len(article['columns']['left'])} 个元素")
print(f"右栏: {len(article['columns']['right'])} 个元素")

局限性

手写布局可能不准确
非常小的文本区域可能被遗漏
复杂的嵌套布局具有挑战性
批量处理建议使用GPU
多语言支持因语言而异

安装

pip install surya-ocr

# 用于PDF处理
pip install pdf2image

layout-analyzer

相关 Skills

布局分析技能

概述

使用方法

领域知识

surya 基础

布局元素类型

文本检测

布局分析

阅读顺序检测

带布局的OCR

处理PDF

可视化

最佳实践

常见模式

文档结构提取

表格区域提取

示例

示例1：学术论文分析

示例2：表单字段检测

示例3：多栏文章

局限性

安装

资源

你可能还喜欢

blucli

agent-architecture-audit

triaging-issues

继续探索