Which file is this picture in? Searching PDFs, PowerPoint and Word files by image

Why ordinary search can't find it

Desktop search builds an index of words: file names, document text, metadata. A picture inside a document is stored as an image stream — a JPEG or PNG packed into the file — and nothing in that index describes what it looks like. There is simply nothing for a text search to match against.

Text recognition does not close the gap either. It helps only when the picture itself contains words. A product photo, a logo variant, a site photo or a scanned drawing has to be found by comparing the images themselves.

The free way: extract the images, then compare them

This works on a Mac or a PC, costs nothing, and is the right answer if you have a few dozen documents to check once.

Step 1 — Word, PowerPoint and Excel: unzip them

Modern Office files are ZIP packages; Microsoft's own documentation tells developers to rename a .pptx to .zip to look inside. Copy the file, change the extension, open it, and the embedded pictures are all in one folder:

report.docx   →  word/media/
deck.pptx     →  ppt/media/
prices.xlsx   →  xl/media/

On Windows, rename the copy to .zip and use Extract All. On a Mac, Terminal's unzip reads the file without renaming it:

unzip deck.pptx "ppt/media/*" -d deck-images

Older .doc, .ppt and .xls files are not ZIP archives. Open them in Office and save them in the current format first.

Step 2 — PDFs: pdfimages

pdfimages ships with Poppler, the open-source PDF library. It copies out the images embedded in the PDF rather than rendering the pages — JPEGs byte for byte, most other images as PNG — and the -p option puts the page number in every file name, which is what lets you trace a match back to a page. On a Mac, install it with Homebrew; on Windows, prebuilt binaries are published at poppler-windows and the commands are the same.

brew install poppler
pdfimages -list catalog.pdf
mkdir catalog-images
pdfimages -all -p catalog.pdf catalog-images/catalog

Pictures with transparency come out as two files, the image and a grey mask; the mask can be ignored. If you already pay for Adobe Acrobat, its Convert panel has an Export all Images option that does the same job, one PDF at a time.

Step 3 — compare against the picture you have

Put your reference picture in with the extracted folders and run a similarity scan: dupeGuru's Picture mode groups it with its matches, and digiKam can search for images similar to it once the folders are added to its collection. Name each extracted folder after its document, and a match tells you the file — plus the page, for PDFs.

These free tools are good at one specific kind of match: the same image, resized or recompressed. If the picture you have is a different photograph of the same product, they will not pair it. Why that is, and what does pair it, is covered in duplicate finders vs. semantic search.

Your options compared

What ImgSeek reads inside documents — and what it doesn't

You point ImgSeek at folders. It indexes the images in them and the images embedded in the documents in them, where they are — nothing is moved, and nothing is uploaded. Each result names the source document and the position inside it:

• PDF — embedded images, page by page: "Page 14, image 2".
• PowerPoint (.pptx) — pictures on every slide, including pictures inside grouped shapes: "Slide 3, image 1".
• Excel (.xlsx and legacy .xls) — pictures on every sheet, labelled with the sheet name.
• Word (.docx) — embedded pictures, numbered "Image 1", "Image 2" and so on. Word documents have no fixed pages, so there is no page number to report.

Not supported: legacy .doc and .ppt files (save them as .docx or .pptx first); Keynote, Pages and Google Docs files; charts and diagrams drawn as vector shapes; and HEIC or RAW image files sitting in the same folders — ImgSeek reads JPEG, PNG, WebP, BMP, GIF and TIFF.

For common kinds of objects, such as chairs, cups or clocks, it also detects the main subject of your reference picture and searches with that, so a shared white background or slide template is less likely to drive the matches. Subject detection only knows a fixed set of common object types: anything else, headphones for example, is missed or given the wrong label. In that case search with the whole image, which still works.

When ImgSeek is the wrong tool

• A few documents, one time — the free method above is enough. Paying for software to check twenty files is poor value.
• You are looking for words, not pictures — your system search already indexes document text, and does it better.
• Your documents live in Keynote, Pages, Google Docs, or a cloud workspace that isn't synced to your disk — ImgSeek only reads local .pdf, .docx, .pptx, .xlsx and .xls files.
• You have an Intel Mac or use Linux — the Mac build is Apple Silicon only, and there is no Linux build.
• You can't install unnotarized apps — the Mac build is not yet notarized by Apple, so the first launch needs a manual approval (steps here). Some company-managed Macs block that outright. ImgSeek is also closed source.
• You need a definitive answer, not a shortlist — results are ranked by similarity score, and someone still has to open the document and confirm.
• You need to single out one kind of object the detector doesn't know — subject detection covers common object types only, so headphones or a specific product part won't be picked out of a busy picture. Searching with the whole image still works.