PDF OCR — 使用 PDF Oxide 从扫描版 PDF 中提取文本

通过内置 OCR 从扫描版 PDF 中提取文本。自 v0.3.27 起，OCR 通过统一的 FFI 层（pdf_ocr_engine_create、pdf_ocr_page_needs_ocr、pdf_ocr_extract_text）向所有语言绑定开放，包括 Python、Node.js、Go、C# 和 Rust。

Python

from pdf_oxide import PdfDocument

doc = PdfDocument("scanned.pdf")
text = doc.extract_text_ocr(0)
print(text)

Node.js

const { PdfDocument, OcrEngine } = require("pdf-oxide");

const doc = new PdfDocument("scanned.pdf");
const ocr = new OcrEngine();
if (ocr.pageNeedsOcr(doc, 0)) {
  console.log(ocr.extractText(doc, 0));
}
ocr.close();
doc.close();

Go

import pdfoxide "github.com/yfedoseev/pdf_oxide/go"

doc, _ := pdfoxide.Open("scanned.pdf")
defer doc.Close()

ocr, _ := pdfoxide.NewOcrEngine()
defer ocr.Close()

if ocr.NeedsOcr(doc, 0) {
    text, _ := ocr.ExtractTextWithOcr(doc, 0)
    fmt.Println(text)
}

C#

using PdfOxide.Core;
using PdfOxide.Ocr;

using var doc = PdfDocument.Open("scanned.pdf");
using var ocr = new OcrEngine();

if (ocr.PageNeedsOcr(doc, 0))
{
    Console.WriteLine(ocr.ExtractText(doc, 0));
}

Rust

use pdf_oxide::PdfDocument;
use pdf_oxide::ocr::{OcrEngine, OcrConfig, OcrExtractOptions, extract_text_with_ocr};

let mut doc = PdfDocument::open("scanned.pdf")?;
let config = OcrConfig::default();
let engine = OcrEngine::new("models/det.onnx", "models/rec.onnx", "models/dict.txt", config)?;
let options = OcrExtractOptions::default();
let text = extract_text_with_ocr(&mut doc, 0, Some(&engine), options)?;
println!("{text}");

Java

import fyi.oxide.pdf.PdfDocument;

try (PdfDocument doc = PdfDocument.open("scanned.pdf")) {
    // Lean-tier bindings have no OCR engine handle — extractTextAuto
    // routes scanned pages through OCR automatically (graceful fallback).
    String text = doc.extractTextAuto(0);
    System.out.println(text);
}

PHP

<?php
use PdfOxide\PdfDocument;

$doc = PdfDocument::open("scanned.pdf");
// No OCR engine handle in the lean tier — extractTextAuto routes
// scanned pages through OCR automatically (graceful fallback).
echo $doc->extractTextAuto(0);

Ruby

require "pdf_oxide"

doc = PdfOxide::PdfDocument.open("scanned.pdf")
# No OCR engine handle in the lean tier — extract_text_auto routes
# scanned pages through OCR automatically (graceful fallback).
puts doc.extract_text_auto(0)

C++

#include <pdf_oxide/pdf_oxide.hpp>

auto doc = pdf_oxide::Document::open("scanned.pdf");
auto engine = pdf_oxide::OcrEngine::create("det.onnx", "rec.onnx", "dict.txt");

if (doc.ocr_page_needs_ocr(0)) {
    std::string text = doc.ocr_extract_text(0, &engine);
    std::cout << text << "\n";
}

Swift

import PdfOxide

let doc = try Document.open("scanned.pdf")
let engine = try OcrEngine.create(
    detModelPath: "det.onnx", recModelPath: "rec.onnx", dictPath: "dict.txt")

if try doc.ocrPageNeedsOcr(0) {
    let text = try doc.ocrExtractText(0, engine: engine)
    print(text)
}

Kotlin

import fyi.oxide.pdf.PdfDocument

PdfDocument.open("scanned.pdf").use { doc ->
    // Lean-tier bindings have no OCR engine handle — extractTextAuto
    // routes scanned pages through OCR automatically (graceful fallback).
    println(doc.extractTextAuto(0))
}

Dart

import 'package:pdf_oxide/pdf_oxide.dart';

final doc = PdfDocument.open('scanned.pdf');
final engine = OcrEngine.create('det.onnx', 'rec.onnx', 'dict.txt');

if (doc.pageNeedsOcr(0)) {
  print(doc.ocrExtractText(0, engine));
}
engine.close();
doc.close();

R

library(pdfoxide)

doc    <- pdf_open("scanned.pdf")
engine <- pdf_ocr_engine_create("det.onnx", "rec.onnx", "dict.txt")

if (pdf_ocr_page_needs_ocr(doc, 0)) {
  text <- pdf_ocr_extract_text(doc, 0, engine)
  cat(text)
}

Julia

using PdfOxide

doc    = open_document("scanned.pdf")
engine = ocr_engine_create("det.onnx", "rec.onnx", "dict.txt")

if page_needs_ocr(doc, 0)
    text = ocr_extract_text(doc, 0, engine)
    println(text)
end

Zig

const pdf_oxide = @import("pdf_oxide");
const a = std.heap.page_allocator;

var doc = try pdf_oxide.Document.open("scanned.pdf");
defer doc.deinit();
var engine = try pdf_oxide.OcrEngine.create("det.onnx", "rec.onnx", "dict.txt");
defer engine.deinit();

if (try doc.ocrPageNeedsOcr(0)) {
    const text = try doc.ocrExtractText(a, 0, engine);
    defer a.free(text);
    std.debug.print("{s}\n", .{text});
}

Scala

import fyi.oxide.pdf.PdfDocument

val doc = PdfDocument.open("scanned.pdf")
// Lean-tier bindings have no OCR engine handle — extractTextAuto
// routes scanned pages through OCR automatically (graceful fallback).
println(doc.extractTextAuto(0))
doc.close()

Clojure

(require '[pdf-oxide.core :as pdf])

(with-open [doc (pdf/open "scanned.pdf")]
  ;; Lean-tier bindings have no OCR engine handle — the AutoExtractor
  ;; routes scanned pages through OCR automatically (graceful fallback).
  (let [ax (pdf/auto-extractor doc)]
    (println (pdf/auto-text ax))))

Objective-C

#import "POXPdfOxide.h"
NSError *err = nil;

POXDocument *doc = [POXDocument openPath:@"scanned.pdf" error:&err];
POXOcrEngine *engine = [POXOcrEngine createWithDetModelPath:@"det.onnx"
                                              recModelPath:@"rec.onnx"
                                                  dictPath:@"dict.txt"
                                                     error:&err];

if ([doc pageNeedsOcr:0 error:&err]) {
    NSString *text = [doc ocrExtractText:0 engine:engine error:&err];
    NSLog(@"%@", text);
}

Elixir

{:ok, doc} = PdfOxide.open("scanned.pdf")
{:ok, engine} = PdfOxide.ocr_engine("det.onnx", "rec.onnx", "dict.txt")

case PdfOxide.ocr_page_needs_ocr(doc, 0) do
  {:ok, true} ->
    {:ok, text} = PdfOxide.ocr_extract_text(doc, 0, engine)
    IO.puts(text)

  _ ->
    :ok
end

PDF Oxide 通过 ONNX Runtime 内置了 PaddleOCR——无需安装 Tesseract，无系统依赖，无子进程调用。OCR 引擎直接在进程内运行，支持 PP-OCRv3、PP-OCRv4 和 PP-OCRv5 模型系列。

注意： WebAssembly 不支持 OCR（需要原生 ONNX Runtime）。Go / Node.js / C# / Rust 需要开启 ocr feature 进行编译。Python wheel 默认已启用 OCR。

无需 Tesseract 的 Python PDF OCR

大多数 Python PDF OCR 方案都需要将 Tesseract 作为系统依赖安装——在不同操作系统和 CI 环境中配置繁琐。PDF Oxide 将 PaddleOCR 模型直接打包到 Python wheel 中：

• 零系统依赖 — pip install pdf_oxide 一行搞定
• 无子进程调用 — OCR 通过 ONNX Runtime 在原生进程内运行
• 三个模型系列 — PP-OCRv3、PP-OCRv4、PP-OCRv5
• 自动页面检测 — 自动识别扫描页与文本页

对比：PDF Oxide OCR vs PyMuPDF + Tesseract

	PDF Oxide	PyMuPDF + Tesseract
安装	pip install pdf_oxide	pip install pymupdf + 系统级 Tesseract
OCR 引擎	PaddleOCR（ONNX）	Tesseract（子进程）
配置复杂度	一行命令	各 OS 单独安装 Tesseract
CI/Docker	无需额外配置	需要 apt-get install tesseract-ocr
模型内置	是（打包在 wheel 中）	否（需单独下载）

安装

Python

pip install pdf_oxide

OCR 模型已内置于 wheel，无需额外下载。

Rust

[dependencies]
pdf_oxide = { version = "0.3", features = ["ocr"] }

Go

go build -tags ocr ./...

Node.js

npm install pdf-oxide --build-from-source -- --features ocr

C#

NuGet 包在默认的 Linux / macOS / Windows 二进制文件中已启用 OCR，无需额外配置。

何时使用 OCR

大多数 PDF 包含嵌入文本，extract_text() 每页仅需 0.8ms。以下情况才需要使用 OCR：

• 扫描文档 — 纸质文档扫描成 PDF
• 纯图像 PDF — 由照片或截图生成的 PDF
• 文本以图像形式存储的 PDF — 部分生成工具会将文本光栅化
• 混合页面 — 同时包含原生文本和扫描图像区域的页面

---

PP-OCR 模型版本

PDF Oxide 支持三代 PaddleOCR 模型。默认配置适用于 PP-OCRv3 和 PP-OCRv4。PP-OCRv5 服务端模型需要使用不同的缩放策略。

PP-OCRv3 / PP-OCRv4（默认）

移动端优化模型，将图像缩小以适应最大边长限制。适合大多数文档场景。

• 检测模型：DBNet++（轻量级）
• 识别模型：SVTR
• 缩放策略：MaxSide — 将最长边缩放至 960px
• 适用场景：标准文档、移动端/边缘部署

Python

from pdf_oxide import OcrConfig, OcrEngine

# Default config works with v3/v4 models
config = OcrConfig()
engine = OcrEngine("det_v4.onnx", "rec_v4.onnx", "dict.txt", config)

Rust

use pdf_oxide::ocr::{OcrConfig, OcrEngine};

// Default config: MaxSide { max_side: 960 }
let config = OcrConfig::default();
let engine = OcrEngine::new("det_v4.onnx", "rec_v4.onnx", "dict.txt", config)?;

PP-OCRv5（服务端）

服务端级别模型，在必要时将图像放大以保持高分辨率。对密集文字或小字体文档的识别精度显著提升。

• 检测模型：DBNet++（服务端版，体积更大）
• 识别模型：SVTR-v5
• 缩放策略：MinSide — 确保最短边不低于 64px，上限为 4000px
• 适用场景：高精度提取、服务端环境、密集文字

Python

from pdf_oxide import OcrConfig, OcrEngine

# v5 config: high-resolution input for server models
config = OcrConfig(use_v5=True)
engine = OcrEngine("det_v5.onnx", "rec_v5.onnx", "dict_v5.txt", config)

Rust

use pdf_oxide::ocr::{OcrConfig, OcrEngine};

// v5 config: MinSide { min_side: 64, max_side_limit: 4000 }
let config = OcrConfig::v5();
let engine = OcrEngine::new("det_v5.onnx", "rec_v5.onnx", "dict_v5.txt", config)?;

模型对比

特性	PP-OCRv3/v4	PP-OCRv5
缩放策略	MaxSide（缩小至 960px）	MinSide（放大，上限 4000px）
输入分辨率	较低（速度快）	较高（精度高）
检测模型大小	约 3 MB	约 12 MB
识别模型大小	约 12 MB	约 25 MB
适用场景	移动端、边缘、标准文档	服务端、密集文字、小字体
OcrConfig	OcrConfig() / OcrConfig::default()	OcrConfig(use_v5=True) / OcrConfig::v5()

---

页面类型检测

PDF Oxide 会自动对页面分类以判断是否需要 OCR。extract_text_ocr() 内部已处理此逻辑，但你也可以手动检测页面类型。

自动检测扫描页

Python

from pdf_oxide import PdfDocument

doc = PdfDocument("mixed.pdf")

for i in range(doc.page_count()):
    text = doc.extract_text(i)
    if len(text.strip()) < 50:
        # Likely scanned — use OCR
        text = doc.extract_text_ocr(i)
        print(f"Page {i + 1} (OCR): {text[:100]}...")
    else:
        print(f"Page {i + 1} (text): {text[:100]}...")

Rust

use pdf_oxide::PdfDocument;
use pdf_oxide::ocr::{detect_page_type, PageType, OcrEngine, OcrConfig, OcrExtractOptions, extract_text_with_ocr};

let mut doc = PdfDocument::open("mixed.pdf")?;
let engine = OcrEngine::new("det.onnx", "rec.onnx", "dict.txt", OcrConfig::default())?;

for i in 0..doc.page_count() {
    let page_type = detect_page_type(&mut doc, i)?;
    match page_type {
        PageType::NativeText => {
            let text = doc.extract_text(i)?;
            println!("Page {} (native): {}...", i + 1, &text[..100.min(text.len())]);
        }
        PageType::ScannedPage => {
            let text = extract_text_with_ocr(&mut doc, i, Some(&engine), OcrExtractOptions::default())?;
            println!("Page {} (OCR): {}...", i + 1, &text[..100.min(text.len())]);
        }
        PageType::HybridPage => {
            // Has both native text and scanned images — merges both sources
            let text = extract_text_with_ocr(&mut doc, i, Some(&engine), OcrExtractOptions::default())?;
            println!("Page {} (hybrid): {}...", i + 1, &text[..100.min(text.len())]);
        }
    }
}

Java

import fyi.oxide.pdf.PdfDocument;

try (PdfDocument doc = PdfDocument.open("mixed.pdf")) {
    for (int i = 0; i < doc.pageCount(); i++) {
        // extractTextAuto classifies each page and routes scanned
        // pages through OCR automatically (graceful fallback).
        System.out.printf("Page %d: %s%n", i + 1, doc.extractTextAuto(i));
    }
}

PHP

<?php
use PdfOxide\PdfDocument;

$doc = PdfDocument::open("mixed.pdf");
for ($i = 0; $i < $doc->pageCount(); $i++) {
    // extractTextAuto classifies each page and routes scanned
    // pages through OCR automatically (graceful fallback).
    printf("Page %d: %s\n", $i + 1, $doc->extractTextAuto($i));
}

Ruby

require "pdf_oxide"

doc = PdfOxide::PdfDocument.open("mixed.pdf")
doc.page_count.times do |i|
  # extract_text_auto classifies each page and routes scanned
  # pages through OCR automatically (graceful fallback).
  puts "Page #{i + 1}: #{doc.extract_text_auto(i)}"
end

C++

#include <pdf_oxide/pdf_oxide.hpp>

auto doc = pdf_oxide::Document::open("mixed.pdf");
auto engine = pdf_oxide::OcrEngine::create("det.onnx", "rec.onnx", "dict.txt");

for (int i = 0; i < doc.page_count(); ++i) {
    std::string text = doc.ocr_page_needs_ocr(i)
        ? doc.ocr_extract_text(i, &engine)   // scanned / hybrid → OCR
        : doc.extract_text(i);               // native text
    std::cout << "Page " << (i + 1) << ": " << text << "\n";
}

Swift

import PdfOxide

let doc = try Document.open("mixed.pdf")
let engine = try OcrEngine.create(
    detModelPath: "det.onnx", recModelPath: "rec.onnx", dictPath: "dict.txt")

for i in 0..<(try doc.pageCount()) {
    let text = try doc.ocrPageNeedsOcr(i)
        ? doc.ocrExtractText(i, engine: engine)   // scanned / hybrid → OCR
        : doc.extractText(i)                      // native text
    print("Page \(i + 1): \(text)")
}

Kotlin

import fyi.oxide.pdf.PdfDocument

PdfDocument.open("mixed.pdf").use { doc ->
    for (i in 0 until doc.pageCount()) {
        // extractTextAuto classifies each page and routes scanned
        // pages through OCR automatically (graceful fallback).
        println("Page ${i + 1}: ${doc.extractTextAuto(i)}")
    }
}

Dart

import 'package:pdf_oxide/pdf_oxide.dart';

final doc = PdfDocument.open('mixed.pdf');
final engine = OcrEngine.create('det.onnx', 'rec.onnx', 'dict.txt');

for (var i = 0; i < doc.pageCount; i++) {
  final text = doc.pageNeedsOcr(i)
      ? doc.ocrExtractText(i, engine)   // scanned / hybrid → OCR
      : doc.extractText(i);             // native text
  print('Page ${i + 1}: $text');
}
engine.close();
doc.close();

R

library(pdfoxide)

doc    <- pdf_open("mixed.pdf")
engine <- pdf_ocr_engine_create("det.onnx", "rec.onnx", "dict.txt")

for (i in seq_len(pdf_page_count(doc)) - 1) {
  text <- if (pdf_ocr_page_needs_ocr(doc, i)) {
    pdf_ocr_extract_text(doc, i, engine)   # scanned / hybrid -> OCR
  } else {
    pdf_extract_text(doc, i)               # native text
  }
  cat(sprintf("Page %d: %s\n", i + 1, text))
}

Julia

using PdfOxide

doc    = open_document("mixed.pdf")
engine = ocr_engine_create("det.onnx", "rec.onnx", "dict.txt")

for i in 0:(page_count(doc) - 1)
    text = page_needs_ocr(doc, i) ?
        ocr_extract_text(doc, i, engine) :   # scanned / hybrid -> OCR
        extract_text(doc, i)                 # native text
    println("Page $(i + 1): $text")
end

Zig

const pdf_oxide = @import("pdf_oxide");
const a = std.heap.page_allocator;

var doc = try pdf_oxide.Document.open("mixed.pdf");
defer doc.deinit();
var engine = try pdf_oxide.OcrEngine.create("det.onnx", "rec.onnx", "dict.txt");
defer engine.deinit();

var i: i32 = 0;
const n = try doc.pageCount();
while (i < n) : (i += 1) {
    const text = if (try doc.ocrPageNeedsOcr(i))
        try doc.ocrExtractText(a, i, engine)   // scanned / hybrid → OCR
    else
        try doc.extractText(a, i);             // native text
    defer a.free(text);
    std.debug.print("Page {d}: {s}\n", .{ i + 1, text });
}

Scala

import fyi.oxide.pdf.PdfDocument

val doc = PdfDocument.open("mixed.pdf")
for (i <- 0 until doc.pageCount) {
  // extractTextAuto classifies each page and routes scanned
  // pages through OCR automatically (graceful fallback).
  println(s"Page ${i + 1}: ${doc.extractTextAuto(i)}")
}
doc.close()

Clojure

(require '[pdf-oxide.core :as pdf])

(with-open [doc (pdf/open "mixed.pdf")]
  ;; The AutoExtractor classifies each page and routes scanned pages
  ;; through OCR automatically (graceful fallback).
  (println (pdf/auto-text (pdf/auto-extractor doc))))

Objective-C

#import "POXPdfOxide.h"
NSError *err = nil;

POXDocument *doc = [POXDocument openPath:@"mixed.pdf" error:&err];
POXOcrEngine *engine = [POXOcrEngine createWithDetModelPath:@"det.onnx"
                                              recModelPath:@"rec.onnx"
                                                  dictPath:@"dict.txt"
                                                     error:&err];

NSInteger n = [doc pageCountError:&err];
for (NSInteger i = 0; i < n; i++) {
    NSString *text = [doc pageNeedsOcr:i error:&err]
        ? [doc ocrExtractText:i engine:engine error:&err]   // scanned / hybrid → OCR
        : [doc extractText:i error:&err];                   // native text
    NSLog(@"Page %ld: %@", (long)(i + 1), text);
}

Elixir

{:ok, doc} = PdfOxide.open("mixed.pdf")
{:ok, engine} = PdfOxide.ocr_engine("det.onnx", "rec.onnx", "dict.txt")
{:ok, n} = PdfOxide.page_count(doc)

for i <- 0..(n - 1) do
  {:ok, text} =
    case PdfOxide.ocr_page_needs_ocr(doc, i) do
      {:ok, true} -> PdfOxide.ocr_extract_text(doc, i, engine)   # scanned / hybrid -> OCR
      _ -> PdfOxide.extract_text(doc, i)                         # native text
    end

  IO.puts("Page #{i + 1}: #{text}")
end

PageType 枚举值（Rust）

枚举值	说明
NativeText	页面包含嵌入文本 — 无需 OCR
ScannedPage	页面为纯扫描（大图像，无文本或极少文本） — 执行完整 OCR
HybridPage	页面同时包含原生文本和扫描图像 — 合并原生文本与 OCR 结果

needs_ocr() 辅助函数对 ScannedPage 和 HybridPage 均返回 true：

use pdf_oxide::ocr::needs_ocr;

if needs_ocr(&mut doc, 0)? {
    let text = extract_text_with_ocr(&mut doc, 0, Some(&engine), OcrExtractOptions::default())?;
}

---

工作原理

1. PDF Oxide 在内部将页面渲染为图像（300 DPI）
2. 根据检测策略对图像进行缩放（v3/v4 使用 MaxSide，v5 使用 MinSide）
3. DBNet++ 文本检测器以四边形边界框定位文本区域
4. SVTR 文本识别器读取每个检测区域中的字符
5. 按阅读顺序排序后将结果组装成文本
6. 混合页面中，OCR 文本与原生文本合并输出

整个流水线通过 ONNX Runtime 在进程内运行，无外部二进制文件、无子进程调用、无临时文件。

---

OCR 配置

Python

from pdf_oxide import OcrConfig, OcrEngine

# Default (v3/v4)
config = OcrConfig()

# PP-OCRv5 server models
config = OcrConfig(use_v5=True)

# Custom thresholds
config = OcrConfig(
    det_threshold=0.5,    # Detection confidence (0.0-1.0)
    box_threshold=0.7,    # Box confidence (0.0-1.0)
    rec_threshold=0.6,    # Recognition confidence (0.0-1.0)
    num_threads=8,        # ONNX Runtime threads
    max_candidates=500,   # Max text regions
)

# v5 with custom thresholds
config = OcrConfig(use_v5=True, det_threshold=0.4, num_threads=8)

engine = OcrEngine("det.onnx", "rec.onnx", "dict.txt", config)

Rust

use pdf_oxide::ocr::{OcrConfig, OcrConfigBuilder, DetResizeStrategy};

// Default (v3/v4): MaxSide { max_side: 960 }
let config = OcrConfig::default();

// PP-OCRv5: MinSide { min_side: 64, max_side_limit: 4000 }
let config = OcrConfig::v5();

// Custom builder
let config = OcrConfig::builder()
    .det_threshold(0.5)
    .box_threshold(0.7)
    .rec_threshold(0.6)
    .num_threads(8)
    .max_candidates(500)
    .detect_styles(true)        // Enable style detection from OCR geometry
    .build();

// Custom resize strategy
let config = OcrConfig::builder()
    .det_resize_strategy(DetResizeStrategy::MinSide {
        min_side: 128,
        max_side_limit: 6000,
    })
    .build();

DetResizeStrategy（Rust）

控制检测模型运行前输入图像的缩放方式。

枚举值	字段	说明
MaxSide	max_side: u32（默认：960）	缩小使最长边不超过 max_side。PP-OCRv3/v4 的默认值。
MinSide	min_side: u32（默认：64），max_side_limit: u32（默认：4000）	放大使最短边不低于 min_side，上限为 max_side_limit。PP-OCRv5 的默认值。

OcrConfig 字段

字段	类型	默认值	说明
det_threshold	f32	0.3	检测概率阈值
box_threshold	f32	0.6	边界框置信度阈值
rec_threshold	f32	0.5	识别置信度阈值
det_max_side	u32	960	图像最大尺寸（v3/v4 兼容）
det_resize_strategy	DetResizeStrategy	MaxSide { 960 }	图像缩放策略
rec_target_height	u32	48	识别裁剪的目标高度
num_threads	usize	4	ONNX Runtime 推理线程数
unclip_ratio	f32	1.5	边界框扩展比率
max_candidates	usize	1000	最大文本区域检测数量
detect_styles	bool	true	从 OCR 几何信息推断字体样式
det_model_path	Option<PathBuf>	None	自定义检测模型路径
rec_model_path	Option<PathBuf>	None	自定义识别模型路径
dict_path	Option<PathBuf>	None	自定义字符词典路径

自定义模型

使用自己的 ONNX 模型替代内置模型：

Rust

use pdf_oxide::ocr::OcrConfig;

let config = OcrConfig::builder()
    .det_model_path("models/custom_det.onnx")
    .rec_model_path("models/custom_rec.onnx")
    .dict_path("models/custom_dict.txt")
    .build();

样式检测

启用 detect_styles（默认开启）后，PDF Oxide 会从 OCR 几何信息（文字大小、间距、位置）推断字体样式（粗体、标题级别），从而提升扫描页 Markdown 转换的输出质量。

let config = OcrConfig::builder()
    .detect_styles(true)    // Infer styles from text geometry
    .build();

---

OCR vs Tesseract

特性	PDF Oxide OCR	Tesseract（通过 PyMuPDF）
安装	pip install pdf_oxide	系统包 + pytesseract
系统依赖	无	需要 Tesseract 二进制文件
运行时	ONNX（进程内）	子进程调用
模型版本	PP-OCRv3、v4、v5	Tesseract LSTM
语言支持	多语言	需要语言包
配置复杂度	零	中等
检测模型	DBNet++	Tesseract 内置
识别模型	SVTR / SVTR-v5	Tesseract LSTM
高分辨率支持	MinSide 策略（v5）	DPI 设置
页面类型检测	自动（原生/扫描/混合）	手动

---

自定义 DPI

控制将 PDF 页面转换为图像时的渲染分辨率：

Python

from pdf_oxide import PdfDocument

doc = PdfDocument("scanned.pdf")

# Default is 300 DPI — good balance of accuracy and speed
text = doc.extract_text_ocr(0)

# Higher DPI for better accuracy on fine print
text = doc.extract_text_ocr(0)  # DPI configured via OcrExtractOptions in Rust

Rust

use pdf_oxide::ocr::OcrExtractOptions;

// Higher DPI = better accuracy but slower
let options = OcrExtractOptions::default().with_dpi(300.0);

// Lower DPI = faster but less accurate
let options = OcrExtractOptions::default().with_dpi(150.0);

---

OCR 输出结构（Rust）

OcrEngine::ocr_image() 方法返回带有每个文字区域置信度分数的详细结果：

use pdf_oxide::ocr::OcrEngine;

let engine = OcrEngine::new("det.onnx", "rec.onnx", "dict.txt", Default::default())?;
let output = engine.ocr_image(&image)?;

// Full text in reading order
println!("{}", output.text_in_reading_order());

// Per-span details
for span in &output.spans {
    println!("Text: '{}' (confidence: {:.2})", span.text, span.confidence);
    println!("  Bounding box: {:?}", span.bounding_rect());
    println!("  Per-char confidence: {:?}", span.char_confidences);
}

// Overall confidence
println!("Total confidence: {:.2}", output.total_confidence);

OcrOutput 字段

字段 / 方法	类型	说明
spans	Vec<OcrSpan>	所有已识别的文本区域
total_confidence	f32	所有文字区域的平均置信度
text()	String	所有文本以空格连接的字符串
text_in_reading_order()	String	按位置排序的文本（从上到下、从左到右）

OcrSpan 字段

字段	类型	说明
text	String	已识别文本
polygon	[[f32; 2]; 4]	四边形边界框（4 个角点）
confidence	f32	整体置信度（0.0–1.0）
char_confidences	Vec<f32>	逐字符置信度分数

---

批量 OCR 处理

处理一个目录中的所有扫描版 PDF：

Python

from pdf_oxide import PdfDocument, PdfError
from pathlib import Path

pdf_dir = Path("scans/")
output_dir = Path("text-output/")
output_dir.mkdir(exist_ok=True)

for pdf_path in pdf_dir.glob("*.pdf"):
    try:
        doc = PdfDocument(str(pdf_path))
        pages = []
        for i in range(doc.page_count()):
            text = doc.extract_text(i)
            if len(text.strip()) < 50:
                text = doc.extract_text_ocr(i)
            pages.append(text)

        out_path = output_dir / pdf_path.with_suffix(".txt").name
        out_path.write_text("\n\n".join(pages), encoding="utf-8")
    except PdfError as e:
        print(f"Error: {pdf_path.name}: {e}")

Rust

use pdf_oxide::PdfDocument;
use pdf_oxide::ocr::{OcrEngine, OcrConfig, OcrExtractOptions, extract_text_with_ocr, needs_ocr};
use std::fs;
use std::path::Path;

let engine = OcrEngine::new("det.onnx", "rec.onnx", "dict.txt", OcrConfig::default())?;
let options = OcrExtractOptions::default();

for entry in fs::read_dir("scans/")? {
    let path = entry?.path();
    if path.extension().map_or(false, |e| e == "pdf") {
        let mut doc = PdfDocument::open(path.to_str().unwrap())?;
        let mut all_text = String::new();
        for i in 0..doc.page_count() {
            let text = if needs_ocr(&mut doc, i)? {
                extract_text_with_ocr(&mut doc, i, Some(&engine), options.clone())?
            } else {
                doc.extract_text(i)?
            };
            all_text.push_str(&text);
            all_text.push_str("\n\n");
        }
        let out_path = Path::new("text-output/")
            .join(path.file_stem().unwrap())
            .with_extension("txt");
        fs::write(out_path, &all_text)?;
    }
}

并行 OCR（Python）

from pdf_oxide import PdfDocument
from multiprocessing import Pool
from pathlib import Path

def ocr_pdf(pdf_path: str) -> dict:
    doc = PdfDocument(pdf_path)
    text = ""
    for i in range(doc.page_count()):
        text += doc.extract_text_ocr(i) + "\n"
    return {"file": pdf_path, "text": text}

pdf_files = [str(p) for p in Path("scans/").glob("*.pdf")]

with Pool(4) as pool:
    results = pool.map(ocr_pdf, pdf_files)

---

OCR 转 Markdown

将扫描页面转换为 Markdown：

Python

from pdf_oxide import PdfDocument

doc = PdfDocument("scanned-report.pdf")

for i in range(doc.page_count()):
    md = doc.to_markdown(i, detect_headings=True)
    if len(md.strip()) < 50:
        # Scanned page — OCR then format
        text = doc.extract_text_ocr(i)
        md = text  # OCR output is plain text
    print(f"--- Page {i + 1} ---")
    print(md)

Rust

use pdf_oxide::PdfDocument;
use pdf_oxide::ocr::{OcrEngine, OcrConfig, OcrExtractOptions, needs_ocr, extract_text_with_ocr};

let mut doc = PdfDocument::open("scanned-report.pdf")?;
let engine = OcrEngine::new("det.onnx", "rec.onnx", "dict.txt", OcrConfig::default())?;

for i in 0..doc.page_count() {
    let text = if needs_ocr(&mut doc, i)? {
        extract_text_with_ocr(&mut doc, i, Some(&engine), OcrExtractOptions::default())?
    } else {
        doc.to_markdown(i, &Default::default())?
    };
    println!("--- Page {} ---\n{}", i + 1, text);
}

Java

import fyi.oxide.pdf.PdfDocument;

try (PdfDocument doc = PdfDocument.open("scanned-report.pdf")) {
    for (int i = 0; i < doc.pageCount(); i++) {
        String md = doc.toMarkdown(i);
        if (md.strip().length() < 50) {
            // Scanned page — auto-routing returns OCR text (plain).
            md = doc.extractTextAuto(i);
        }
        System.out.printf("--- Page %d ---%n%s%n", i + 1, md);
    }
}

PHP

<?php
use PdfOxide\PdfDocument;

$doc = PdfDocument::open("scanned-report.pdf");
for ($i = 0; $i < $doc->pageCount(); $i++) {
    $md = $doc->toMarkdown($i);
    if (strlen(trim($md)) < 50) {
        // Scanned page — auto-routing returns OCR text (plain).
        $md = $doc->extractTextAuto($i);
    }
    printf("--- Page %d ---\n%s\n", $i + 1, $md);
}

Ruby

require "pdf_oxide"

doc = PdfOxide::PdfDocument.open("scanned-report.pdf")
doc.page_count.times do |i|
  md = doc.to_markdown(i)
  if md.strip.length < 50
    # Scanned page — auto-routing returns OCR text (plain).
    md = doc.extract_text_auto(i)
  end
  puts "--- Page #{i + 1} ---\n#{md}"
end

C++

#include <pdf_oxide/pdf_oxide.hpp>

auto doc = pdf_oxide::Document::open("scanned-report.pdf");
auto engine = pdf_oxide::OcrEngine::create("det.onnx", "rec.onnx", "dict.txt");

for (int i = 0; i < doc.page_count(); ++i) {
    std::string text = doc.ocr_page_needs_ocr(i)
        ? doc.ocr_extract_text(i, &engine)   // scanned / hybrid → OCR
        : doc.to_markdown(i);                // native → Markdown
    std::cout << "--- Page " << (i + 1) << " ---\n" << text << "\n";
}

Swift

import PdfOxide

let doc = try Document.open("scanned-report.pdf")
let engine = try OcrEngine.create(
    detModelPath: "det.onnx", recModelPath: "rec.onnx", dictPath: "dict.txt")

for i in 0..<(try doc.pageCount()) {
    let text = try doc.ocrPageNeedsOcr(i)
        ? doc.ocrExtractText(i, engine: engine)   // scanned / hybrid → OCR
        : doc.toMarkdown(i)                       // native → Markdown
    print("--- Page \(i + 1) ---\n\(text)")
}

Kotlin

import fyi.oxide.pdf.PdfDocument

PdfDocument.open("scanned-report.pdf").use { doc ->
    for (i in 0 until doc.pageCount()) {
        var md = doc.toMarkdown(i)
        if (md.trim().length < 50) {
            // Scanned page — auto-routing returns OCR text (plain).
            md = doc.extractTextAuto(i)
        }
        println("--- Page ${i + 1} ---\n$md")
    }
}

Dart

import 'package:pdf_oxide/pdf_oxide.dart';

final doc = PdfDocument.open('scanned-report.pdf');
final engine = OcrEngine.create('det.onnx', 'rec.onnx', 'dict.txt');

for (var i = 0; i < doc.pageCount; i++) {
  final text = doc.pageNeedsOcr(i)
      ? doc.ocrExtractText(i, engine)   // scanned / hybrid → OCR
      : doc.toMarkdown(i);              // native → Markdown
  print('--- Page ${i + 1} ---\n$text');
}
engine.close();
doc.close();

R

library(pdfoxide)

doc    <- pdf_open("scanned-report.pdf")
engine <- pdf_ocr_engine_create("det.onnx", "rec.onnx", "dict.txt")

for (i in seq_len(pdf_page_count(doc)) - 1) {
  text <- if (pdf_ocr_page_needs_ocr(doc, i)) {
    pdf_ocr_extract_text(doc, i, engine)   # scanned / hybrid -> OCR
  } else {
    pdf_to_markdown(doc, i)                # native -> Markdown
  }
  cat(sprintf("--- Page %d ---\n%s\n", i + 1, text))
}

Julia

using PdfOxide

doc    = open_document("scanned-report.pdf")
engine = ocr_engine_create("det.onnx", "rec.onnx", "dict.txt")

for i in 0:(page_count(doc) - 1)
    text = page_needs_ocr(doc, i) ?
        ocr_extract_text(doc, i, engine) :   # scanned / hybrid -> OCR
        to_markdown(doc, i)                  # native -> Markdown
    println("--- Page $(i + 1) ---\n$text")
end

Zig

const pdf_oxide = @import("pdf_oxide");
const a = std.heap.page_allocator;

var doc = try pdf_oxide.Document.open("scanned-report.pdf");
defer doc.deinit();
var engine = try pdf_oxide.OcrEngine.create("det.onnx", "rec.onnx", "dict.txt");
defer engine.deinit();

var i: i32 = 0;
const n = try doc.pageCount();
while (i < n) : (i += 1) {
    const text = if (try doc.ocrPageNeedsOcr(i))
        try doc.ocrExtractText(a, i, engine)   // scanned / hybrid → OCR
    else
        try doc.toMarkdown(a, i);              // native → Markdown
    defer a.free(text);
    std.debug.print("--- Page {d} ---\n{s}\n", .{ i + 1, text });
}

Scala

import fyi.oxide.pdf.PdfDocument

val doc = PdfDocument.open("scanned-report.pdf")
for (i <- 0 until doc.pageCount) {
  var md = doc.toMarkdown(i)
  if (md.trim.length < 50) {
    // Scanned page — auto-routing returns OCR text (plain).
    md = doc.extractTextAuto(i)
  }
  println(s"--- Page ${i + 1} ---\n$md")
}
doc.close()

Clojure

(require '[pdf-oxide.core :as pdf])

(with-open [doc (pdf/open "scanned-report.pdf")]
  ;; The AutoExtractor routes scanned pages through OCR automatically.
  (println (pdf/auto-text (pdf/auto-extractor doc))))

Objective-C

#import "POXPdfOxide.h"
NSError *err = nil;

POXDocument *doc = [POXDocument openPath:@"scanned-report.pdf" error:&err];
POXOcrEngine *engine = [POXOcrEngine createWithDetModelPath:@"det.onnx"
                                              recModelPath:@"rec.onnx"
                                                  dictPath:@"dict.txt"
                                                     error:&err];

NSInteger n = [doc pageCountError:&err];
for (NSInteger i = 0; i < n; i++) {
    NSString *text = [doc pageNeedsOcr:i error:&err]
        ? [doc ocrExtractText:i engine:engine error:&err]   // scanned / hybrid → OCR
        : [doc toMarkdown:i error:&err];                    // native → Markdown
    NSLog(@"--- Page %ld ---\n%@", (long)(i + 1), text);
}

Elixir

{:ok, doc} = PdfOxide.open("scanned-report.pdf")
{:ok, engine} = PdfOxide.ocr_engine("det.onnx", "rec.onnx", "dict.txt")
{:ok, n} = PdfOxide.page_count(doc)

for i <- 0..(n - 1) do
  {:ok, text} =
    case PdfOxide.ocr_page_needs_ocr(doc, i) do
      {:ok, true} -> PdfOxide.ocr_extract_text(doc, i, engine)   # scanned / hybrid -> OCR
      _ -> PdfOxide.to_markdown(doc, i)                          # native -> Markdown
    end

  IO.puts("--- Page #{i + 1} ---\n#{text}")
end

---

性能考量

OCR 比文本提取慢得多：

操作	典型速度
文本提取	每页 0.8ms
OCR（v3/v4）	每页 200–1,000ms
OCR（v5 服务端）	每页 500–2,000ms

OCR 速度取决于页面复杂度、图像分辨率、文字密度和模型版本。PP-OCRv5 速度较慢但精度更高。处理大批量文件时，建议使用并行处理（参见上方"批量 OCR 处理"）。

---

从字节加载模型（Rust）

use pdf_oxide::ocr::{OcrEngine, OcrConfig};

let det_bytes = std::fs::read("models/det.onnx")?;
let rec_bytes = std::fs::read("models/rec.onnx")?;
let dict = std::fs::read_to_string("models/dict.txt")?;

let engine = OcrEngine::from_bytes(&det_bytes, &rec_bytes, &dict, OcrConfig::default())?;

---

相关页面

• 文本提取 — 标准文本提取
• Markdown 转换 — 带标题检测的 Markdown
• 页面渲染 — 将页面渲染为图像（OCR 内部使用）
• 批量处理 — 并行处理模式
• 从 PDF 提取文本 — 文本提取指南