import EventEmitter from 'events'; import sharp from 'sharp'; import PDFJS from 'pdfjs-dist/es5/build/pdf'; import NodeCanvasFactory from './NodeCanvasFactory'; /** @typedef generatedPoint @type {Object} @property {number} x The x coordinate. @property {number} y The y coordinate. */ const DEFAULT_PARSER_OPTIONS = { /** * The minimum distance between two lines on the y-axis to consider them part * of separate paragraphs. This option only affects the `parse` method. * * Default: `25`. */ paragraphThreshold: 25, /** * The minimum distance between two lines on the y-axis to consider them part * of the same line. PDFs usually suffer from issues with floating point * numbers. This value is used to give a little room for error. You shouldn't * have to change this value unless you're dealing with PDFs generated with * OCR or other odd PDFs. * * Default: `1`. */ lineThreshold: 1, /** * Scaling applied to the PDF before extrating images. Higher value results in * greater image resolution, but quadratically increases rendering times. * * Default: `2` */ imageScale: 2, /** * Controls whether or not to extract images. Image extraction requires * rendering of each page, which might take a long time depending on the size * of the PDF, configured `imageScale` and underlying hardware. If you don't * need to extract images, setting this option to `false` is recommended. * * Default: `true` */ extractImages: true, /** * Controls whether or not to ignore empty text elements. Text elements are * considered empty if their text content contains nothing by whitespace. * * Default: `true` */ ignoreEmptyText: true, /** * Controls whether or not to join paragraphs. Enabling this option will join * each line that's not separated by a non-text element (paragraph break or * image) which will effectively make each line contain a paragraph. Paragraph * breaks will be omitted from the final output. This option only affects the * `parse` method. * * Default: `false` */ joinParagraphs: false, /** * Controls what format the image is exported as. Defaults to 'png'. Passed * directly to Sharp: https://sharp.pixelplumbing.com/api-output#toformat */ imageOutputFormat: 'png', }; /** * Parses a PDF file into its components and exposes them through either events * or promises. */ export default class SimplePDFParser extends EventEmitter { /** * * @param {Buffer} fileBuffer Buffer containing PDF file * @param {DEFAULT_PARSER_OPTIONS} [options] */ constructor(fileBuffer, options = {}) { super(); /** * Contains the PDF passed in through the constructor * @type {Buffer} */ this.fileBuffer = fileBuffer; /** * Options for the parser * @type {DEFAULT_PARSER_OPTIONS} */ this.options = { ...DEFAULT_PARSER_OPTIONS, ...options, }; } /** * Parses the PDF file passed in the constructor. Returns a promise that * resolves with an array of all elements. */ async parse() { const pages = await this.parseRaw(); const elements = []; for (let i = 0; i < pages.length; i += 1) { for (let j = 0; j < pages[i].textElements.length; j += 1) { const element = pages[i].textElements[j]; elements.push({ type: 'text', pageIndex: i, y: element.y, items: element.items, }); } if (this.options.extractImages) { for (let j = 0; j < pages[i].imageElements.length; j += 1) { const element = pages[i].imageElements[j]; elements.push({ type: 'image', pageIndex: i, y: element.y, imageBuffer: element.imageBuffer, }); } } } const elementsSorted = elements .slice() .sort((a, b) => (a.pageIndex - b.pageIndex !== 0 ? a.pageIndex - b.pageIndex : a.y - b.y)); const elementsJoined = []; for (let i = 0; i < elementsSorted.length; i += 1) { const elem = elementsSorted[i]; const { y: _, ...elemClean } = elem; const lastElem = elementsSorted[i - 1]; let deltaY = null; if (lastElem && lastElem.type === 'text') { deltaY = elem.y - lastElem.y; deltaY = deltaY > 0 ? deltaY : -deltaY; } if (deltaY !== null && deltaY > this.options.paragraphThreshold) { if (!this.options.joinParagraphs) { elementsJoined.push({ type: 'paragraphBreak', }); } elementsJoined.push(elemClean); } else { const lastJoined = elementsJoined[elementsJoined.length - 1]; if (this.options.joinParagraphs && lastJoined && lastJoined.type === 'text') { if (lastJoined.items.length > 0) { const lastItem = lastJoined.items[lastJoined.items.length - 1]; lastItem.text = `${lastItem.text} `; } lastJoined.items = [ ...lastJoined.items, ...elemClean.items, ]; } else { elementsJoined.push(elemClean); } } } return elementsJoined; } /** * Parses the PDF file passed in the constructor. Returns a promise that * resolves with an object containing text and image elements separately. */ async parseRaw() { const loadingTask = PDFJS.getDocument(new Uint8Array(this.fileBuffer)); const pdf = await loadingTask.promise; const pagePromises = []; for (let i = 1; i <= pdf.numPages; i += 1) { pagePromises.push(this.parsePage(pdf, i)); } const results = await Promise.all(pagePromises); this.emit('done'); return results .sort((a, b) => a.index - b.index) .map((page) => ({ textElements: page.textElements, imageElements: page.imageElements, })); } async parsePage(pdf, pageIndex) { const page = await pdf.getPage(pageIndex); const textData = await page.getTextContent(); const textElements = []; let imageElements = null; for (let i = 0; i < textData.items.length; i += 1) { const item = textData.items[i]; let shouldParse = true; if (this.options.ignoreEmptyText) { const textContent = item.str.replace(/\u200B/, ''); // Remove zero-width spaces shouldParse = textContent.trim().length !== 0; } if (shouldParse) { const lastElem = textElements[textElements.length - 1]; const y = page.view[3] - item.transform[5]; let deltaY = null; if (lastElem) { deltaY = y - lastElem.y; deltaY = deltaY > 0 ? deltaY : -deltaY; } if (deltaY !== null && deltaY < this.options.lineThreshold) { lastElem.items.push({ text: item.str, font: item.fontName, }); } else { textElements.push({ x: item.transform[4], y, items: [{ text: item.str, font: item.fontName, }], }); } } } if (this.options.extractImages) { imageElements = await this.extractPageImages(page); } const result = { index: pageIndex, textElements, imageElements, }; this.emit('page', result); return result; } async extractPageImages(page) { const ops = await page.getOperatorList(); /** * `[width, skew, skew, height, x, y]` */ let lastMatrix; const imageRects = []; const pageWidth = page.view[2]; const pageHeight = page.view[3]; for (let i = 0; i < ops.fnArray.length; i += 1) { const fn = ops.fnArray[i]; const args = ops.argsArray[i]; switch (fn) { case PDFJS.OPS.transform: { lastMatrix = args; break; } case PDFJS.OPS.paintImageXObject: { // For the matrix see https://github.com/mozilla/pdf.js/issues/10498 // Also, this comment // https://github.com/mozilla/pdf.js/issues/5643#issuecomment-496648719 // links to a PDF reference file explaining the translation array // format. You can find it on page 142: // https://via.hypothes.is/https://www.adobe.com/content/dam/acom/en/devnet/pdf/pdfs/pdf_reference_archives/PDFReference.pdf#annotations:SVudloF5EemLBgPm0gmY3Q // Or this actually seems to be the original link: // https://www.adobe.com/content/dam/acom/en/devnet/pdf/pdfs/pdf_reference_archives/PDFReference.pdf const width = lastMatrix[0]; const height = lastMatrix[3]; const heightAbs = height > 0 ? height : -height; // x is always relative to the left edge const x = lastMatrix[4]; // y might be relative to the top or bottom edge depending on previous // transformations. We later calculate `top` later to unify it. // y is always relative to the bottom edge of the image. const y = lastMatrix[5]; // If height is negative, y is relative to the top edge. // If height is positive, y is relative to the bottom edge. // Subtract the absolute height. const top = (height < 0 ? y : pageHeight - y) - heightAbs; // Clamp the width and height so that they don't overflow the page. const widthClamped = (x + width) > pageWidth ? (pageWidth - x) : width; const heightClamped = (top + heightAbs) > pageHeight ? (pageHeight - top) : heightAbs; imageRects.push({ x, y: top, width: widthClamped, height: heightClamped, }); break; } default: // do nothing } } let images = []; if (imageRects.length > 0) { const pageImageBuffer = await this.renderPage(page); const pageSharp = sharp(pageImageBuffer); images = await Promise.all(imageRects.map(async (elem) => ({ ...elem, imageBuffer: await pageSharp .clone() .extract({ left: Math.round(elem.x * this.options.imageScale), top: Math.round(elem.y * this.options.imageScale), width: Math.round(elem.width * this.options.imageScale), height: Math.round(elem.height * this.options.imageScale), }) .toFormat(this.options.imageOutputFormat) .toBuffer(), }))); } return images; } async renderPage(page) { const viewport = page.getViewport({ scale: this.options.imageScale, }); const canvasFactory = new NodeCanvasFactory(); const canvasAndContext = canvasFactory.create( viewport.width, viewport.height, ); const renderContext = { canvasContext: canvasAndContext.context, viewport, canvasFactory, }; const renderTask = page.render(renderContext); await renderTask.promise; return canvasAndContext.canvas.toBuffer(); } }