Image to Text (OCR)

Image OCR runs Tesseract.js v5 — a WebAssembly port of the open-source Tesseract engine originally built at HP and now maintained by Google. The library is loaded from a JSDelivr CDN bundle on first use; the WebAssembly binary plus the language data file for your selected language are cached by the browser, so subsequent runs start almost instantly.

Tesseract uses an LSTM-based recurrent neural network as its core recognition model, with a separate page layout analyser that finds text blocks, lines, and word boundaries before character recognition begins. The combination handles multi-column text, mixed font sizes, and rotated lines reasonably well — better than many older template-matching OCR engines.

Recognition runs on a Web Worker so the UI stays responsive. Your image is decoded into a Canvas, the pixel buffer is handed to the Tesseract worker, and progress is reported back as a percentage from layout analysis through final recognition. A typical 1080-pixel screenshot finishes in two to five seconds on a desktop, ten to twenty seconds on a phone for a full A4 scan.

Thirteen languages are supported in this build: English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Simplified Chinese, Japanese, Korean, Arabic, and Hindi. Each language has its own data file (the English file is roughly 5 MB; CJK languages are larger). Select the right language before running OCR — using English on a French document drops accuracy noticeably because the language model influences character disambiguation.

Output includes the recognised text and a confidence score per line and per word. Confidence above 80 percent is usually accurate; below 70 percent the result needs a human pass. For best accuracy, pre-process your image: crop to just the text region, increase contrast in Image Enhancer, and rotate the image upright before running OCR.

Handwriting support is limited. Tesseract was designed for printed text and handles neat block printing acceptably; it struggles with cursive, ornate scripts, and overlapping handwritten lines. For genuine handwriting recognition, a model trained specifically on handwriting is required.

All processing is local. The image is read into a Canvas, the pixel buffer goes to a Worker that runs Tesseract.js, and the recognised text is held in React state. There is no network call carrying image bytes — only the one-time CDN fetch for the engine and language data, which the browser then caches.