A scan of a page is, to a computer, just a grid of colored dots — it “contains” text the way a photograph of a piano contains music. OCR is the technology that reads those dots and reconstructs the actual characters.
The pipeline, in plain English
1. Preprocessing: the image is straightened, cleaned of noise, and separated into dark ink versus light background. 2. Layout analysis: the engine finds the blocks — columns, paragraphs, lines — so words get read in the right order. 3. Recognition: a neural network trained on millions of text samples reads each line’s shapes as character sequences, using language statistics to arbitrate ambiguous shapes (is that “rn” or “m”?). 4. Reconstruction: recognized text is reassembled with positions, which is how our OCR to Word tool rebuilds headings, paragraphs, and alignment rather than emitting a text blob.
Why OCR fails, when it fails
Low resolution starves the network of shape detail — 300 DPI scans succeed where phone thumbnails fail. Skew, shadows, and busy backgrounds confuse layout analysis. Decorative fonts fall outside training data. Handwriting is a fundamentally different (and harder) problem than print.
Using it well
The practical workflow lives in the scanned-PDF guide; modern browser OCR runs fully locally, which the no-upload guide explains matters for confidential scans. If your PDF already has selectable text, skip OCR — direct extraction is exact.