Extract Text from a Mixed PDF: Digital Pages, Scans and Missing OCR

A mixed PDF can contain normal digital text on one page and a scanned picture of text on another. It can mix both on the same sheet: a selectable heading above a photographed receipt, for example. Finding one digital text object does not prove that every visible word on that sheet is available as text. The default extraction path reads native PDF text first and recognizes pages with no extracted text. That is efficient for ordinary digital documents and fully scanned sheets. It can miss words in an image on a page that already contains a digital heading. For that case, enable OCR every PDF page and run a fresh extraction. The full-page route recognizes the rendered appearance instead of treating a little native text as evidence that the whole page is covered. This is a mode choice rather than an automatic accuracy upgrade. A native digital paragraph usually extracts more faithfully than a newly recognized image of that paragraph. Full-page OCR can introduce errors, and it takes longer. Use the mode matching the missing information, then compare the transcript with the pages you need.

The public example above contains a digital first page and a second page with a selectable heading plus a printed-English image. The image includes a made-up invoice reference DEMO-104 and a total of USD 42.00. It is demonstration content, not a merchant invoice or financial record. The accompanying PNG contains the same scan text so you can test image recognition separately. Run the PDF through the default extraction mode. Check whether the first-page marker is present. Then look for DEMO-104 and the amount from the image. A transcript containing the second page's heading but not the reference illustrates why a successful extraction can still be incomplete. Enable OCR every PDF page, generate a fresh transcript and compare it with the sample. Check the reference character by character, particularly the hyphen and digits. OCR can confuse zero with O and one with I. This example makes checks reproducible; it is not an accuracy benchmark for all scans. A clear synthetic image is simpler than a faded photograph, handwritten note or dense multi-column statement.

This tool produces a transcript that you can review, edit, copy or download as a text file. It does not add an invisible OCR layer to the original PDF. Searching the transcript and searching the PDF itself are different outcomes. Be clear about which one the receiving system or colleague needs. Plain text is useful when you need a quotation, reference number or editable words for another application. It does not preserve every visual relationship from the source. A line break can change; a table may turn into a sequence of fragments. Do not infer that a total belongs to the preceding label without looking back at the scan. If your goal is to open the PDF in a reader and use Find inside that file, choose software that explicitly exports a searchable PDF with a text layer. Downloading this transcript does not accomplish that. Keep the source alongside the reviewed text, and tell a recipient when a passage was recovered with OCR. For records requiring an exact representation of the original, a transcript is an aid, not a replacement.

Start with a readable image. Look for small text, skewed lines, shadows, faint ink and a busy background. If you have the original paper, a clearer scan often helps more than repeatedly running the same low-quality file. Use enough resolution for the smallest characters without creating an unnecessarily huge image. Crop irrelevant surroundings when appropriate, but retain information that the document must keep. Straighten the page and use even lighting for a phone photograph. Check the result at a normal reading zoom rather than trusting a thumbnail. Compression that makes punctuation or thin strokes disappear can damage recognition even when the page still looks roughly recognizable. The current model is for printed English. Handwriting, other scripts and complex layouts may be inaccurate. English letters inside another-language document do not make the entire page an English scan. For a difficult table, verify row and column associations manually or use structured-extraction software. More processing time is not proof of better understanding. The useful test is whether recovered words, numbers and their associations agree with the source.

Recognition happens on your device, so its speed and available memory matter. A large multipage scan can run slowly on a phone even when the initial page loads quickly. The first recognition downloads the worker, compatible WebAssembly engine and English model from this website. That software loading is separate from uploading your PDF. This release has explicit file, page and image-pixel limits as well as bounded worker initialization and recognition. If a limit or timeout is reached, split the PDF into smaller groups and process the relevant pages. Keep a note of each group's original physical page range so separate transcripts can be put back in the right order. Use Cancel extraction if the job is no longer needed. After a failure, the interface should show a useful error and allow a fresh attempt instead of leaving the last transcript looking like a new result. If an engine download was blocked, check the connection or browser restrictions and retry the public example. Do not disable security software indiscriminately or wait indefinitely for a damaged source.

Read the transcript while keeping the original open beside it. Pay particular attention to dates, identifiers, currency values, email addresses and proper names. A sentence can look fluent while one critical digit is wrong. For the practice scan, compare DEMO-104 and USD 42.00 with the actual image rather than accepting a plausible substitute. The search box filters displayed text; it does not validate a match. A missing result can mean recognition failed, punctuation changed or the source contains a different spelling. Search a shorter distinctive fragment and inspect the relevant page. Do not conclude that a record lacks a word solely because transcript search returned nothing. Make corrections in the editable transcript and download the reviewed version. Copying it to another program, sending it by email or saving it to a synchronized folder creates separate data handling outside this website. Protect sensitive transcripts like the original document. For an important quotation, record the physical source page so another person can verify it. OCR provides candidate text; responsible use still requires comparison with the source.

Selected document bytes, the editable transcript and any PDF password are not submitted to a recognition API in this release. Native extraction and image recognition run in browser code and workers. Ordinary page requests, local engine downloads and optional consented visit analytics are separate network activity. Inspect the boundary with a safe public sample in your browser's network panel. Engine requests should retrieve software or model resources rather than send the selected file as a request body. A local design does not promise secure erasure of memory or protection from a malicious extension. Downloads remain on your device, and your operating system may have its own backup or synchronization behavior. In a support report, include the browser, extraction mode, approximate page count and error message. Prefer a public practice file or a made-up example reproducing the behavior. Do not send a private scan, transcript or password to support for convenience. Explain whether native text is missing, the image engine fails to initialize or recognized words are wrong: those are different problems with different fixes. The public example provides a shared starting point for that discussion.