Why OCR Misses Equations, Footnotes, Tables, and Diagrams

A searchable PDF made from screenshots holds two things: the page images and an invisible text layer produced by OCR. The images are exact. The text layer is a best guess, and on textbook pages it is weakest in the parts students search most: equations, footnotes, tables and labelled diagrams.

This post explains why those parts fail, which capture settings help, and how to check a PDF before you depend on its search.

Why textbook pages are difficult

OCR finds blocks of text in an image, splits them into lines and words, and matches each character shape against a model of a written language. It works best on plain paragraphs: one column, regular spacing, body-size type and good contrast.

Textbooks break those assumptions on almost every page. They mix two columns, sidebars, captions, numbered equations, tables and figures. When OCR fails, it does not show an error. The text layer simply holds wrong characters or nothing for that region.

Equations are two-dimensional

A line of prose reads left to right. An equation also reads up and down: fractions stack terms, exponents and subscripts sit off the baseline, and summation limits sit above and below a symbol. Greek letters and operators fall outside the alphabet most OCR models expect.

The usual result is a run of letters and digits in the wrong order, or the equation is skipped. Searching for a variable name may still find the page. Copying the equation into notes rarely works.

Footnotes are small and out of order

Footnotes and captions are set smaller than body text. At screen resolution a small letter may be only a few pixels tall, and similar shapes such as rn and m, or 1 and l, become hard to separate.

Footnotes also sit at the bottom of the page, far from the marker in the text. OCR reads them in position order, so a footnote can appear in the text layer between two unrelated paragraphs.

Tables need row and column structure

OCR outputs lines of text. A table needs rows and columns. Cell borders confuse line detection, and tight columns can merge into one string.

Numbers usually survive as characters. Their row and column order often does not, so a copied table may list values in the wrong cells.

Diagram labels are not the diagram

Labels inside a figure are short, rotated or placed over shapes and lines. OCR may skip them, merge neighbours or read a chart axis as characters. The figure itself, its arrows, curves and structure, never becomes text.

For study, treat a figure as an image with a caption. Search for words in the caption, not in the figure.

A practical verification workflow

1. Search for body text

   Open the PDF in Preview and search for a phrase from the middle of a paragraph. It should match on the right page.

2. Search for a footnote word

   Pick an uncommon word from a footnote or caption. A miss shows that small type is the weak point.

3. Copy an equation

   Drag across an equation and paste it into a text editor. Gaps and scrambled symbols show how much of the math the text layer holds.

4. Copy a table row

   Paste one row and compare it with the image. Check that values stay in their columns.

5. Check a two-column page

   Copy a paragraph from the left column. Words from the right column in the paste mean the reading order is mixed.

For a single figure or table, run Extract Text from Image on a cropped capture to see what can be read, then correct it by hand.

Improve recognition before capture

Most OCR problems start with how many pixels each character gets. You control that before export.

• Use a Retina display. A Retina screenshot stores about twice the pixels per side of the logical window size.
• Zoom in. Raise the reader's zoom so body text is clearly larger than the default. Footnotes improve the most.
• Capture one column. Draw a Custom Area around one column, or use a single-column view if the reader offers one.
• Wait for rendering. A longer Auto Capture interval prevents half-loaded text.
• Use a light theme. Dark backgrounds and coloured highlight bands lower contrast.

Temposhot captures what is on screen at the resolution the display provides. It cannot recover detail the reader did not render.

Do not OCR around a copying restriction

OCR in Temposhot is meant for pages you are allowed to save, such as your own material or titles whose license permits personal copies.

A review note to keep with the file

Add a short note next to each exported PDF, or as the first page of a Document:

• Source edition and page range.
• Capture date and zoom level.
• Known OCR gaps, for example "equations on pages 14 to 18 not searchable".
• Pages checked against the original.

The note tells you, months later, which pages to trust for search and which to read as images.

FAQ

Why does search find the page but not the exact phrase

Line breaks, hyphenation and mixed reading order split phrases in the text layer. Search for one uncommon word instead.

Does a higher JPG quality improve OCR

Slightly. Resolution matters far more. Zoom and a Retina display give larger gains than export quality.

Can I fix the text layer by hand

Not inside the PDF in a practical way. Type the important equations and table values into your notes and keep the PDF for reference.

Working with pages you are allowed to save?

Capture them at a higher zoom and run the checks above on the first export. The save screenshots as PDF page shows the full workflow.