Skip to content

armgit / papra

Public Imported from GitHub
0

The minimalistic document archiving platform.

feat(pdf-extractor): add page rendering fallback for pdf without text nor image (#948)

Corentin Thomasset725eaff4b033
diff --git a/.changeset/two-dodos-invent.md b/.changeset/two-dodos-invent.md
new file mode 100644
index 00000000..268a5552
--- /dev/null
+++ b/.changeset/two-dodos-invent.md
@@ -0,0 +1,6 @@
+---
+"@papra/lecture": minor
+"@papra/app": minor
+---
+
+When extracting text from PDF documents, if neither text nor images suitable for OCR are found, the pages are rendered as images and processed with OCR. Adding support for vectorized text.
\ No newline at end of file
diff --git a/apps/papra-server/src/modules/documents/documents.services.ts b/apps/papra-server/src/modules/documents/documents.services.ts
index 727ad489..1731ebbd 100644
--- a/apps/papra-server/src/modules/documents/documents.services.ts
+++ b/apps/papra-server/src/modules/documents/documents.services.ts
@@ -18,12 +18,14 @@ export async function extractDocumentText({
ocrLanguages?: string[];
logger?: Logger;
}) {
+ const startTime = Date.now();
const { textContent, error, extractorName, extractorType } = await extractTextFromFile({ file, config: { tesseract: { languages: ocrLanguages } }, logger });
+ const durationMs = Date.now() - startTime;
if (error) {
- logger.error({ error, extractorName, extractorType }, 'Error while extracting text from document');
+ logger.error({ error, extractorName, extractorType, durationMs }, 'Error while extracting text from document');
} else {
- logger.info({ extractorName, extractorType }, 'Text extracted from document');
+ logger.info({ extractorName, extractorType, durationMs }, 'Text extracted from document');
}
return {
diff --git a/packages/lecture/fixtures/017-vectorized-text-pdf/017.expected.txt b/packages/lecture/fixtures/017-vectorized-text-pdf/017.expected.txt
new file mode 100644
index 00000000..80fe60e3
--- /dev/null
+++ b/packages/lecture/fixtures/017-vectorized-text-pdf/017.expected.txt
@@ -0,0 +1,12 @@
+Lorem ipsum
+
+Lorem ipsum dolor sit amet, consectetur
+adipiscing elit. Sapien ante conubia
+vestibulum ultrices quisque nam nascetur
+consectetur. Viverra amet lacinia massa
+donec gravida primis leo tellus. Montes
+nulla sit cras odio penatibus cum aenean
+metus. Per per eros fusce et platea et
+feugiat ullamcorper. Nunc suscipit
+senectus suscipit convallis duis accumsan
+feugiat maecenas.
\ No newline at end of file
diff --git a/packages/lecture/fixtures/017-vectorized-text-pdf/017.input.pdf b/packages/lecture/fixtures/017-vectorized-text-pdf/017.input.pdf
new file mode 100644
index 00000000..f2e55e90
Binary files /dev/null and b/packages/lecture/fixtures/017-vectorized-text-pdf/017.input.pdf differ
diff --git a/packages/lecture/package.json b/packages/lecture/package.json
index 71927736..30717598 100644
--- a/packages/lecture/package.json
+++ b/packages/lecture/package.json
@@ -46,6 +46,7 @@
},
"dependencies": {
"@corentinth/chisels": "catalog:",
+ "@napi-rs/canvas": "^0.1.73",
"fast-xml-parser": "^5.3.0",
"jszip": "^3.10.1",
"rtf-parser": "^1.3.3",
diff --git a/packages/lecture/src/extractors/pdf.extractor.ts b/packages/lecture/src/extractors/pdf.extractor.ts
index d6b182d6..df410417 100644
--- a/packages/lecture/src/extractors/pdf.extractor.ts
+++ b/packages/lecture/src/extractors/pdf.extractor.ts
@@ -1,17 +1,98 @@
+import type { PDFDocumentProxy } from 'unpdf/pdfjs';
+import type { Logger } from '../types';
+import { Buffer } from 'node:buffer';
+import canvas from '@napi-rs/canvas';
import sharp from 'sharp';
-import { extractImages, extractText, getDocumentProxy } from 'unpdf';
+import { extractImages, extractText, getDocumentProxy, renderPageAsImage } from 'unpdf';
import { defineTextExtractor } from '../extractors.models';
import { createTesseractExtractor } from '../tesseract/tesseract.usecases';
+type OcrExtract = (imageBuffer: Buffer) => Promise<string>;
+
+async function extractPdfText({ pdf }: { pdf: PDFDocumentProxy }) {
+ const { text, totalPages: pageCount } = await extractText(pdf, { mergePages: true });
+
+ return {
+ text: text?.trim().length > 0 ? text : undefined,
+ pageCount,
+ };
+}
+
+async function ocrPdfEmbeddedImages({ pdf, pageCount, extract, logger }: { pdf: PDFDocumentProxy; pageCount: number; extract: OcrExtract; logger?: Logger }) {
+ const imageTexts: string[] = [];
+
+ for (let pageIndex = 1; pageIndex <= pageCount; pageIndex++) {
+ const images = await extractImages(pdf, pageIndex);
+ const imageCount = images.length;
+
+ if (imageCount === 0) {
+ logger?.debug({ pageIndex, pageCount }, 'No images found on PDF page for OCR.');
+ continue;
+ }
+
+ logger?.debug({ pageIndex, pageCount, imageCount }, 'Extracted images from PDF page.');
+
+ for (const [imageIndex, image] of images.entries()) {
+ const startTime = Date.now();
+ const imageBuffer = await sharp(image.data, {
+ raw: { width: image.width, height: image.height, channels: image.channels },
+ })
+ .png()
+ .toBuffer();
+
+ const bufferDelay = Date.now() - startTime;
+ logger?.debug({
+ pageIndex,
+ pageCount,
+ imageIndex,
+ imageCount,
+ durationMs: bufferDelay,
+ imageWidth: image.width,
+ imageHeight: image.height,
+ imageSizeBytes: image.data.length,
+ }, 'Converted image to PNG buffer for OCR.');
+
+ const imageText = await extract(imageBuffer);
+ const ocrDelay = Date.now() - startTime - bufferDelay;
+ logger?.debug({ pageIndex, pageCount, imageIndex, imageCount, durationMs: ocrDelay }, 'Extracted text from image using OCR.');
+ imageTexts.push(imageText);
+ }
+ }
+
+ return { imageTexts };
+}
+
+async function ocrPdfRenderedPages({ pdf, pageCount, extract, logger }: { pdf: PDFDocumentProxy; pageCount: number; extract: OcrExtract; logger?: Logger }) {
+ const renderedTexts: string[] = [];
+
+ for (let pageIndex = 1; pageIndex <= pageCount; pageIndex++) {
+ const startTime = Date.now();
+ const pageImage = await renderPageAsImage(pdf, pageIndex, {
+ canvasImport: async () => canvas,
+ scale: 2,
+ });
+
+ const renderDelay = Date.now() - startTime;
+ logger?.debug({ pageIndex, pageCount, durationMs: renderDelay }, 'Rendered PDF page as image for OCR.');
+
+ const pageText = await extract(Buffer.from(pageImage));
+ const ocrDelay = Date.now() - startTime - renderDelay;
+ logger?.debug({ pageIndex, pageCount, durationMs: ocrDelay }, 'Extracted text from rendered page using OCR.');
+ renderedTexts.push(pageText);
+ }
+
+ return { renderedTexts };
+}
+
export const pdfExtractorDefinition = defineTextExtractor({
name: 'pdf',
mimeTypes: ['application/pdf'],
extract: async ({ arrayBuffer, config, logger }) => {
const pdf = await getDocumentProxy(arrayBuffer);
- const { text, totalPages: pageCount } = await extractText(pdf, { mergePages: true });
+ const { text, pageCount } = await extractPdfText({ pdf });
- if (text && text.trim().length > 0) {
+ if (text) {
return {
content: text,
subExtractorsUsed: ['pdf-text'],
@@ -22,53 +103,25 @@ export const pdfExtractorDefinition = defineTextExtractor({
const { extract, extractorType } = await createTesseractExtractor(config.tesseract);
- const imageTexts = [];
- const startOcrTime = Date.now();
-
- for (let pageIndex = 1; pageIndex <= pageCount; pageIndex++) {
- const images = await extractImages(pdf, pageIndex);
- const imageCount = images.length;
-
- if (imageCount === 0) {
- logger?.debug({ pageIndex, pageCount }, 'No images found on PDF page for OCR.');
- continue;
- }
-
- logger?.debug({ pageIndex, pageCount, imageCount }, 'Extracted images from PDF page.');
-
- for (const [imageIndex, image] of images.entries()) {
- const startTime = Date.now();
- const imageBuffer = await sharp(image.data, {
- raw: { width: image.width, height: image.height, channels: image.channels },
- })
- .png()
- .toBuffer();
-
- const bufferDelay = Date.now() - startTime;
- logger?.debug({
- pageIndex,
- pageCount,
- imageIndex,
- imageCount,
- durationMs: bufferDelay,
- imageWidth: image.width,
- imageHeight: image.height,
- imageSizeBytes: image.data.length,
- }, 'Converted image to PNG buffer for OCR.');
-
- const imageText = await extract(imageBuffer);
- const ocrDelay = Date.now() - startTime - bufferDelay;
- logger?.debug({ pageIndex, pageCount, imageIndex, imageCount, durationMs: ocrDelay }, 'Extracted text from image using OCR.');
- imageTexts.push(imageText);
- }
+ const { imageTexts } = await ocrPdfEmbeddedImages({ pdf, pageCount, extract, logger });
+
+ if (imageTexts.length > 0) {
+ logger?.info({ pageCount, imagesProcessedCount: imageTexts.length }, 'Completed OCR on PDF images.');
+
+ return {
+ content: imageTexts.join('\n'),
+ subExtractorsUsed: [extractorType],
+ };
}
- const totalOcrDuration = Date.now() - startOcrTime;
+ logger?.debug({ pageCount }, 'No images found in PDF, falling back to page rendering for OCR.');
+
+ const { renderedTexts } = await ocrPdfRenderedPages({ pdf, pageCount, extract, logger });
- logger?.info({ pageCount, imagesProcessedCount: imageTexts.length, durationMs: totalOcrDuration }, 'Completed OCR on PDF images.');
+ logger?.info({ pageCount, pagesRendered: renderedTexts.length }, 'Completed OCR on rendered PDF pages.');
return {
- content: imageTexts.join('\n'),
+ content: renderedTexts.join('\n'),
subExtractorsUsed: [extractorType],
};
},
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 955ead2f..14c30d1a 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -803,6 +803,9 @@ importers:
'@corentinth/chisels':
specifier: 'catalog:'
version: 2.1.0
+ '@napi-rs/canvas':
+ specifier: ^0.1.73
+ version: 0.1.73
fast-xml-parser:
specifier: ^5.3.0
version: 5.3.3
@@ -13789,17 +13792,6 @@ snapshots:
nanostores: 1.0.1
zod: 4.1.12
- dependencies:
- '@better-auth/utils': 0.3.0
- '@better-fetch/fetch': 1.1.18
- '@standard-schema/spec': 1.0.0
- better-call: 1.1.5([email protected])
- jose: 6.1.0
- kysely: 0.28.8
- nanostores: 1.0.1
- zod: 4.1.12
-
dependencies:
@@ -13813,9 +13805,9 @@ snapshots:
expo-network: 8.0.8([email protected])([email protected])
dependencies:
'@better-auth/utils': 0.3.0
'@better-fetch/fetch': 1.1.18
@@ -15739,7 +15731,6 @@ snapshots:
'@napi-rs/canvas-linux-x64-gnu': 0.1.73
'@napi-rs/canvas-linux-x64-musl': 0.1.73
'@napi-rs/canvas-win32-x64-msvc': 0.1.73
- optional: true
'@napi-rs/[email protected]':
dependencies:
@@ -18531,13 +18522,13 @@ snapshots:
dependencies:
'@better-auth/utils': 0.3.0
'@better-fetch/fetch': 1.1.18
'@noble/ciphers': 2.0.1
'@noble/hashes': 2.0.1
- better-call: 1.1.5([email protected])
+ better-call: 1.1.5([email protected])
defu: 6.1.4
jose: 6.1.0
kysely: 0.28.8