Extract Text with OCR

OCR is a multi-stage process. First, the input image is pre-processed: converted to greyscale, binarised (thresholded to black and white), deskewed (rotation corrected), and de-noised. Next, Tesseract's layout analysis segments the image into text regions, lines, and words. Finally, its LSTM model processes each line of text as a sequence of character probabilities and decodes the most likely character sequence using beam search.

Older OCR engines (including Tesseract 3.x) used a pattern-matching approach — comparing each character image against a library of stored templates. Tesseract 5's LSTM engine instead trains a recurrent neural network on millions of text examples, learning contextual cues at the line level. This dramatically improves accuracy on cursive text, unusual fonts, and degraded documents where individual characters are ambiguous but their context makes them clear.

Before passing an image to Tesseract, we apply several pre-processing steps via ImageMagick: deskewing corrects documents photographed at an angle; binarisation converts colour or grey images to clean black-and-white text; de-noising removes scanner noise and compression artefacts; contrast enhancement makes faint text darker. These steps can recover several percentage points of accuracy on poor-quality scans.

A searchable PDF preserves the original scanned image as the visible layer, with an invisible text layer positioned precisely over each word. Generating this correctly requires mapping Tesseract's word bounding boxes (in image pixels) to PDF coordinates (in points, at the original DPI). We use hOCR output from Tesseract — a structured HTML format with bounding box data — to place each word accurately in the output PDF.

Tesseract supports 100+ languages via separate trained data files. Selecting the correct language pack significantly improves accuracy — using an English model on French text will misidentify accented characters. For documents with mixed languages (common in legal documents, academic papers, or multilingual forms), select a combined language model (e.g. eng+fra) or use the auto-detect mode which runs a language identification pass before recognition.