What Is OCR and How Does On-Device Neural Recognition Make Scanned PDFs Searchable?

Paper documents, scanned receipts, old book archives, and courtroom exhibits often arrive as "image-only" PDFs. You open the file in your reader, press Ctrl+F to look up an invoice number, and receive zero results. You try to drag your cursor across a paragraph to copy a quote, and the entire page simply drags as a static picture.

To transform these inert pictures into dynamic digital assets, document workflows rely on Optical Character Recognition (OCR). In this article, we examine how neural vision models convert pixel arrays into structured text and how the PDF specification embeds an invisible text layer to make scans searchable without changing how they look.

1. Image-Only PDFs vs. Searchable PDFs: What’s the Difference?

From a viewer's perspective, an image-only PDF and a searchable PDF can appear visually identical on screen. Under the hood, their object structures are completely different:

• Image-Only PDF: Contains only an /XObject raster image. There are zero font tables, zero character codes, and zero text stream operators. The document is essentially a digital photocopy.
• Searchable ("Sandwich") PDF: Contains the original scanned bitmap image displayed visually, but paired with a transparent vector text stream overlaid at identical coordinate positions. When you drag your cursor, you select the hidden text layer.

2. The Stages of the Optical Character Recognition Pipeline

Extracting text from raw pixels is a multi-stage machine vision problem:

1. Binarization & Thresholding: The scanner image is converted to grayscale, and an adaptive threshold algorithm (such as Otsu's method) calculates optimal contrast boundaries, separating dark ink from background paper grain and shadows.
2. Deskewing & Page Orientation: If the physical paper was placed crookedly on the scanner glass, geometric line-detection algorithms calculate the tilt angle and rotate the bitmap to align text lines horizontally.
3. Layout & Line Segmentation: The engine identifies page columns, paragraphs, and individual line baselines, preventing text columns from merging together.
4. Neural Character Classification: Deep neural networks (LSTM recurrent neural networks in modern engines like Tesseract) evaluate character glyph shapes and contextual word probabilities to predict the most likely character sequence.
5. Bounding Box Coordinate Generation: For every recognized word and glyph, the engine records precise Cartesian bounding coordinates: [x, y, width, height].

3. How the Invisible Text Layer Is Injected into the PDF

Once character strings and coordinates are extracted, how does the engine make them selectable without covering up the scan? The PDF specification provides a specialized text rendering mode: Rendering Mode 3 (Neither fill nor stroke text).

BT
/F1 10 Tf
3 Tr       % 3 Tr = Invisible text rendering mode!
1 0 0 1 72 650 Tm
(Quarterly Revenue Report) Tj
ET

By setting text mode to 3 Tr, the PDF viewer measures font metrics, maps hitboxes for mouse selection, and indexes words for text search, but renders zero visual ink on the canvas. The original scan bitmap remains 100% visible, while the document behaves exactly like a native digital document.

4. Running OCR Locally with WebAssembly in PDFZento

Historically, running OCR required heavy desktop software or sending files to expensive cloud APIs. In PDFZento OCR PDF, we compile the Tesseract OCR engine directly into WebAssembly. Your browser downloads the compact neural model once, executes image analysis across background Web Workers, and synthesizes the searchable PDF right inside your device RAM with zero cloud upload risk.