OCR review: keep corrections tied to source

Keep the source page close enough to inspect

Scanner Pro switches between Document and Text while keeping a Page 1 label. PDF Scanner App Scanner+ Docs calls its surface Proofread and offers Image and Text views. Both approaches acknowledge that reviewing recognized characters requires access to the original, not just a more polished text block.

Preserve the current page and approximate reading position when switching views. If the document has several similar pages, keep enough identity visible to prevent someone correcting page 2 while inspecting page 3. Where the product supports region mapping, use it to reconnect text with the image; do not imply pixel-level alignment if that mapping is unavailable.

Keep source and result states distinct when one is still loading. A blank image pane should not look like an empty document, and a missing transcript should not imply there was no text. Provide a clear return path after zooming the source so people can resume correction without repeatedly finding the same paragraph.

Name what editing will change

iScan shows a cursor inside extracted text under Page 1, with Language, Copy and Share below. PDF Scanner AI displays an editable text area and keyboard alongside Copy all, To Word, Re-Recognize and Translate. The surrounding commands should make clear whether edits belong to a transcript or to a new document artifact.

Retain corrections as deliberate work. A person may fix one character, join a broken line or restore a missing heading. Do not silently replace their buffer when another recognition result arrives. Make Save, Done or confirmation describe the editing boundary where necessary, especially when the original image remains unchanged.

Give undo a useful scope and preserve a recoverable draft when the keyboard closes. Correction should not require copying the entire result into another app merely to avoid losing it. If the product cannot edit the source document itself, say so naturally rather than allowing the interface to imply that visible pixels have changed.

Make re-recognition a new result, not a silent overwrite

Mobile Scanner App combines Image and OCR views with Edit, Copy, a re-recognition action and Export. Scanner App Scan Documents shows Auto language and a separate Re-recognize command. These controls expose a second attempt, but their data model needs to distinguish retrying recognition from preserving a corrected result.

Before replacing edited text, retain the corrected version or explain the replacement clearly. If a new language choice affects only the next recognition run, show that relationship. Changing a label from Auto to a named language should not make existing text look newly verified when the underlying result has not changed.

Associate recognition jobs with the source-page revision and requested language configuration. A late response for an earlier crop must not replace the current page’s work. If re-recognition fails, keep the prior usable result and identify the failed attempt. An error should not convert a corrected transcript into an empty screen.

Make copy scope and confirmation specific

SCAN ACE places recognized text in a sheet beneath the source invoice, with Copy and Translate and a Copied to Clipboard message. Scan to PDF Converter Scanner combines a source preview and extracted text with a copy confirmation and Share. A toast can confirm delivery to the clipboard, but it does not explain which pages or edits were included.

Differentiate Copy selection, Copy page and Copy all when the product supports those scopes. Keep the action associated with the current corrected version. If an app copies only visible text or only the current page, do not label it Copy all without clarifying the document boundary. Long results need a dependable way to inspect what will leave the app.

Confirm copying only after the platform operation succeeds. A selection change or background result should not alter the copied payload between tap and completion. Treat recognized contact information, account details and private documents as sensitive user content, not as convenient examples to place in analytics or support logs.

Separate recognized text from the format it enters

PDF Scanner Scan Documents App presents Recognize, Copy and Share beneath a text-focused result. PDF Editor shows a simpler OCR page with copy and share icons. A sparse toolbar can be usable, but the destination still needs to explain whether the result is plain text, an editable document or a PDF with another representation of the content.

Do not equate a successful text extraction with preservation of layout. Columns, tables, reading order and paragraph boundaries need separate treatment. A pasted transcript can lose structural meaning even when many characters are correct. Preview the supported output and explain limitations where a structured document is being converted into a simpler form.

Keep translation and recognition separate too. Translating extracted text creates another version with another language, not a correction of the original scan. Preserve access to the source and any user corrections, and identify which version is being shared. A generic Share icon should not silently choose a translated or uncorrected variant.

Store provenance for page and text versions

Represent the source-page revision, recognition result and corrected buffer as related but separate records. Include the relevant language and recognition configuration with the result. Keep a human-edited revision from being overwritten merely because the app refreshes a document list or receives a delayed job completion.

Define correction and regeneration boundaries explicitly. A source crop or rotation can invalidate text-region mapping, while a typo correction need not alter the source image. If the app synchronizes edits, resolve concurrent changes without pretending the newest network response is necessarily the newest user decision. Preserve recoverable work when a conflict needs review.

Document-editing products may also create a separate edited file. Adobe Scan’s editing documentation describes a save-copy boundary for its own editing workflow. That is a useful distinction to explain in your product, not evidence that every OCR screen shown here modifies the source PDF in the same way.

Test correctness at the handoff, not just recognition completion

Prepare a synthetic page with similar characters, a two-column section, a table and one intentionally difficult line. Ask a tester to locate an error, inspect the source, correct it, switch pages and copy the final text. Compare the delivered content with their corrections rather than checking only whether the recognition job returned successfully.

Run another recognition pass after editing, change the selected language and interrupt a job by reopening another document. Verify that the right version remains available and the latest result cannot attach to the wrong page. Test large text, right-to-left material and keyboard focus around long result areas and confirmation actions.

Measure lost corrections, wrong-page exports, repeated recognition attempts and copy-scope mistakes separately from technical job success. Recognition quality needs its own evaluation on representative material. Do not interpret a lower correction count as higher accuracy without checking whether people noticed errors or simply could not find a practical correction path.

Review the source, correction and delivered version

Use these checks on a working build, not only on the polished reference state.

• ✓Every result retains recognizable source-page identity.
• ✓Switching image and text preserves the review context.
• ✓Editing explains whether it changes a transcript or document artifact.
• ✓Re-recognition cannot silently discard human corrections.
• ✓Language changes have a clear effect and scope.
• ✓Copy and export use the intended corrected pages and version.
• ✓Late jobs, failures and synchronization preserve recoverable work.

Connect capture quality with a reliable text handoff

The camera and scanner examples cover producing a usable source image. OCR review starts after capture and needs its own correction model.

Use the sharing and export examples to distinguish the reviewed transcript from the file or clipboard content someone receives.