PDF OCR text corruption: trace the source before preprocessing

Unreadable text from a PDF can start at several different stages: the document's text layer, image extraction, page rendering, recognition, or text export. Changing thresholding cannot repair every one of those faults. This troubleshooting workflow helps you locate the first stage that changes correct information into incorrect information. Start with one representative page, preserve an untouched baseline, and compare the actual image sent to recognition against the page displayed in a PDF viewer.

1. Identify what “corruption” means

First determine whether the wrong characters come from recognition or from extracting existing PDF text. These are different paths, even when they produce similarly unreadable output.

Do not diagnose encoding from appearance alone. An OCR engine can return valid Unicode characters that simply do not match the page. Conversely, correctly recognized text can become unreadable when a file is decoded using the wrong encoding.

If the raw text is correct before application filtering, follow the separate checks for missing OCR characters. There is little value in changing PDF extraction when the information disappears later.

2. Classify each page before choosing a path

A PDF can contain selectable text, scanned page images, vector outlines, or combinations of these. Different pages in the same document can require different handling.

1. Try selecting and copying a complete sentence. Compare its characters and reading order against the displayed page.
2. Check several regions. A selectable heading does not prove that the table beneath it contains usable text.
3. Inspect any extracted text before accepting it. A scanned page may carry an invisible OCR layer containing earlier recognition mistakes.
4. Render one complete page. Keep this image as the visual reference for all subsequent extraction tests.

Use direct text extraction when its content is accurate and its reading order is usable. OCR is unnecessary for reliable existing text.

Use rendered-page OCR when text is absent, outlined, incorrectly mapped, or represented by an unreliable hidden text layer. For mixed pages, choose either a complete rendered-page pass or a region-based workflow that prevents duplicate text.

Record the selected path for each page. A document-level assumption such as “every page is a scan” can silently discard useful text or create duplicate content.

3. Verify page rendering before extracting embedded images

For page-level OCR, start with a proper full-page render rather than collecting every image resource. Rendering reconstructs the page's visible appearance, including placement, rotation, clipping, and text drawn over images.

An embedded image is not necessarily a page. It might be a logo, a background tile, a mask, or one section of a larger scan. Resource enumeration also does not establish reading order or prove that every resource appears visibly on the page.

PDF image streams require interpretation. Depending on the object and extraction interface, the returned bytes might represent a compressed image file or decoded sample data. Width, height, component depth, color space, decode information, and masks may be needed to reconstruct those samples correctly.

Passing arbitrary stream bytes to a general image-file decoder is therefore not a reliable universal extraction method. Use a PDF-aware image extraction interface when you specifically need original image objects. Confirm whether its output is an encoded file or an already reconstructed pixel image.

Perform these tests in order:

• Save the complete rendered page and compare it with the viewer.
• Save each extracted image separately, preserving its dimensions and color mode in a log.
• Identify which visible page region each image represents.
• Check orientation, missing tiles, inverted tones, transparency, and unusual colors.
• Send the verified full-page render through OCR before testing individual objects.

If rendered-page OCR works but extracted-object OCR fails, concentrate on object interpretation and page composition. That comparison is stronger evidence than repeatedly changing recognition engines.

If the rendered page itself differs from the viewer, resolve rendering settings, page bounds, rotation, or document compatibility first.

4. Calculate render size instead of trusting a resolution label

Render density determines how many pixels represent the page. For PDF page dimensions expressed in points, use the standard conversion of 72 points per inch.

For an unrotated page measuring 612 × 792 points, the estimated dimensions at 300 dpi are:

• Width: 612 × 300 / 72 = 2550 pixels.
• Height: 792 × 300 / 72 = 3300 pixels.
• Pixel count: 2550 × 3300 = 8,415,000 pixels.

An uncompressed grayscale buffer at one byte per pixel occupies about 8.4 MB. Three-channel color requires about 25.2 MB, excluding additional buffers and processing overhead.

A practical diagnostic comparison is 200, 300, and 400 dpi on the same difficult region. These are test settings, not guaranteed recognition requirements. Keep language, orientation, and layout settings unchanged during this comparison.

For scanned material, increasing render density cannot restore detail missing from the embedded scan. It mostly interpolates existing samples. For vector text, higher render density can produce genuinely finer rasterized edges.

Inspect saved images at actual pixel scale. A viewer fitted to the window can hide damaged strokes or make a low-resolution page appear smoother than its recognition input.

5. Check language, orientation, and layout before filtering

Once the input image matches the page, verify the recognition configuration. A clear image can still produce nonsense when its writing system or layout does not match the selected recognition mode.

Check that the required language data is installed and actually loaded by the running process. A configured directory is not proof that the intended data was found. Use a known, clean sample containing the expected writing system as a control.

Next, inspect orientation. Page rotation metadata and the orientation of an extracted image are not necessarily equivalent. Correct sideways text before evaluating character recognition, and use a supported vertical-text workflow when the document requires one.

Finally, match segmentation to the region:

• A single text block suits a tightly cropped paragraph.
• Multiple columns require layout handling or separate crops.
• Forms often need fields isolated from rules and neighboring labels.
• Sparse labels need different treatment from a dense page of prose.

Test a paragraph, a table cell, and a short label separately. If those crops work while the complete page fails, investigate layout and reading order before adding stronger image filters.

6. Remove preprocessing until the baseline is understood

Start with the verified render in its original color mode or a straightforward grayscale conversion. Save that baseline before thresholding, inversion, resizing, or morphology.

Then compare a small set of controlled variants:

1. Baseline render without thresholding.
2. Grayscale conversion only.
3. One thresholding method with recorded settings.
4. A targeted cleanup applied only where a demonstrated defect exists.

Do not combine several changes and judge only the final text. You need to know which operation improved or damaged the characters.

Morphological opening can remove thin foreground features along with noise. Its behavior depends on polarity because conventional binary morphology operates on the foreground represented by the high-valued pixels. Verify whether letters or background occupy that foreground before interpreting the effect.

A 3 × 3 kernel spans three image pixels, not a fixed physical feature. If a punctuation mark or character stroke is only two pixels wide, that operation can remove information you need. Dense characters are particularly vulnerable when internal gaps and short strokes are already poorly sampled.

The distinction between point, local, and global preprocessing operations helps explain why the same filter can behave differently across a page.

If resizing or geometric correction is necessary, avoid repeated resampling. Use the guidance on antialiasing during downsampling and image warping to preserve detail rather than introducing another uncontrolled variable.

7. Validate recognition separately from text export

Create a small reference set before tuning. Include at least one clear page, one difficult page, and one mixed-layout page. Manually transcribe representative regions, including punctuation and spaces where they matter.

A useful character-level measure is:

For a reference containing 200 characters, suppose alignment identifies 6 substitutions, 2 deletions, and 2 insertions:

Record how you handle whitespace and Unicode normalization. Changing those rules between runs makes the comparison misleading. Character error rate also does not measure whether table values remain attached to the correct labels.

Use side-by-side OCR inspection grids to keep each crop beside its recognized text. This makes missing strokes, line-order mistakes, and repeated regions easier to locate.

For export, write text using an explicit encoding such as UTF-8. Read the file back with the same encoding and compare the strings before any display formatting. If the strings match but another application displays them incorrectly, investigate that application's decoding or font support.

Preserve raw recognition output separately from cleaned text. Otherwise, later substitutions can conceal the stage where the original error appeared.

8. Consider recapture only when the source lacks detail

Camera replacement is not a remedy for incorrect PDF stream decoding or a broken text layer. Consider recapture only when the underlying scan is blurred, clipped, severely compressed, or too coarsely sampled to retain necessary strokes.

If you can recapture the original document, estimate the sampling requirement before choosing hardware. For an example region measuring 200 × 150 mm, suppose the smallest relevant stroke is 0.20 mm wide. An initial engineering target of three pixels across that stroke gives:

This is a starting sampling budget, not an OCR guarantee. Lens contrast, focus, illumination, motion, and document flatness still affect recognition. Use the camera-resolution calculator to test alternative fields and detail sizes.

For a recapture station, compare area scan cameras through the camera finder. The LS-CMR-0053 and LS-CMR-0166 provide catalog candidates to evaluate against your acquisition requirements, rather than fixes for a PDF software fault.

Check the actual horizontal and vertical pixel counts, document orientation, interface, and shutter requirements before selecting either candidate. Follow the document imaging lens selection guide to evaluate edge sharpness and full-region coverage, then begin the camera configuration workflow if recapture is justified.

9. Confirm the fix before processing the complete document

A successful repair should survive more than the page used during tuning. Run the same settings across your reference set and check these items:

• Page accounting: every intended page is processed once, with no missing or duplicated regions.
• Image fidelity: saved recognition inputs match the displayed pages in orientation and content.
• Text accuracy: difficult characters improve without damaging previously correct regions.
• Reading order: columns, fields, and table entries retain the required relationships.
• Export integrity: saved and reloaded text matches the raw string.
• Repeatability: the same input and recorded settings reproduce the same processing path.

Keep the original PDF, rendered page, recognition input, raw text, and configuration together during diagnosis. For sensitive documents, use an approved processing environment and retain only authorized diagnostic material.

If the fault remains unresolved, prepare an authorized, non-sensitive sample and a stage-by-stage comparison for application support. State whether the first mismatch appears in direct extraction, rendering, image reconstruction, recognition, or export; that evidence determines the next useful test.