how OCR works
OCR does not "read" an image the way a person does: it segments the page, tests character hypotheses against a neural network trained per language, and returns text alongside a confidence score. This tool runs the full Tesseract engine inside the browser via WebAssembly, and the chosen content mode decides how the image gets segmented before recognition.
Single-block segmentation is the fastest option when the text already runs in regular lines, like a scanned document.
Forcing a single line stops the engine from trying to break the text into paragraphs that don't exist.
Without a dictionary the engine stops "correcting" valid characters toward the closest word in the language, which would corrupt a code.
No. Recognition happens in your browser via WebAssembly; the image never leaves your device. Only the OCR engine and the chosen language files are downloaded from a CDN on the first run, and they are cached afterwards.
It works on both, but accuracy drops on a phone photo because of angle, shadow, and focus. Preprocessing (grayscale, 1st-99th percentile contrast stretch, upscaling) compensates for part of that, and auto mode retries with a different segmentation when confidence falls below 55.
Because sparse mode turns the dictionary off on purpose (so it does not "correct" a code toward the closest real word), and without a dictionary the engine decides character by character purely from glyph shape, where 0/O and 1/l/I look nearly identical.
Technically as many as you want, but more than 2 languages at once makes every recognition pass noticeably slower, because the engine tests the text against each checked language dictionary, on top of downloading 11 to 13 MB per pack the first time.
Yes: the Tesseract WebAssembly core and the language pack are downloaded once and cached by the browser, so later passes run locally without needing to fetch anything from the network again.
On the first run the browser downloads the language data (~11 MB in total) from a CDN. After that it is cached.
Use “Sparse/Code” for plates, serial numbers, codes or short strings that are not sentences (like Morse code). In Automatic mode the tool already tries that mode on its own when the first result has low confidence.
Recognition runs entirely in your browser (WebAssembly). The image is never sent to a server; only the OCR engine and language files are downloaded from a CDN on the first run.