Image to Text (OCR)

How this tool actually processes your image

The thumbnail you see in the preview is deliberately downscaled to fit within 800×600 — that keeps the Blazor Server circuit's payload small. But that shrunk preview is only for display: the OCR engine itself runs against your original, full-resolution file, not the thumbnail. A low-res preview on screen doesn't mean lower OCR accuracy.

The engine is Tesseract.js, loaded via a dynamic import() of its ESM build from a CDN rather than a classic <script> tag — script-tag injection doesn't reliably expose the library on window for either v4 or v5 once the page has already booted, so this avoids a class of "Tesseract is not defined" failures. It's also explicitly configured to OCR engine mode 1 — Tesseract's neural-net (LSTM) recognizer only, skipping the older pattern-matching legacy engine entirely. The LSTM engine is the one trained for general real-world text; the legacy engine mostly exists for compatibility with old Tesseract 3 use cases this tool doesn't need.

What actually affects OCR accuracy

• Resolution. Aim for at least 150 DPI for clean digital text, 300 DPI for scanned documents — below that, character strokes start blurring together and the model misreads similar shapes (rn vs m, 0 vs O).
• Contrast. Dark text on a light background works best — Tesseract's training data skews heavily toward this, and low-contrast or color-on-color text degrades fast.
• Rotation and noise. The LSTM model expects roughly horizontal text; heavy rotation, blur, or heavily stylized fonts fall outside what it was trained to recognize confidently.
• Monospace screenshots extract unusually well. Code or terminal screenshots tend to come back nearly perfect — fixed-width glyphs with consistent spacing are close to the clean, structured input the model performs best on.

How to use this tool

1. Click the file input and choose an image (PNG, JPG, WebP, BMP, or TIFF).
2. A preview of the image appears below the input.
3. Click Extract Text. The OCR engine loads on first use (~5 MB from CDN), then processes the image.
4. Copy the extracted text from the output panel.

Privacy: no upload, not even to this site's own server

This is worth being specific about, since most "free OCR" tools online work by uploading your image to a server. This one doesn't — the entire pipeline runs in WebAssembly inside your browser tab, reading the file straight from the local file picker. No image bytes are ever sent anywhere, to this site's servers or anyone else's. A fresh Tesseract worker is created and torn down on every click rather than kept alive in the background, but because the browser caches the ~5 MB engine and language files after the first load, repeat extractions skip the network fetch and only pay the (fast) worker startup cost.

That privacy property is also why this tool is a reasonable choice for screenshots that shouldn't leave your machine — internal dashboards, error messages with environment details, or a terminal screenshot containing a connection string — in a way that uploading to a cloud OCR API genuinely isn't, since that upload step itself is a data exposure regardless of how the provider's retention policy reads.