What We Detect
Image Text (OCR)
Extract and analyze text from images, screenshots, and photos.
QR Codes
Decode QR codes to reveal hidden links and spam content.
Common Catches
- Crypto scam images with wallet addresses
- QR codes linking to phishing sites
- Promotional images with spam text
- Screenshots of spam messages
How text and codes are pulled out of images
Spammers move their payload into pictures precisely to dodge text filters, so Telm reads the images too. Every image runs through OCR — optical character recognition — to extract any text it contains, whether that is a wallet address in a crypto-scam graphic, a promo written over a photo, or a screenshot of a spam message. In parallel, any QR code is decoded to reveal the link hidden inside, often a phishing or malware destination the image never shows in plain text. The extracted text and decoded links are then fed back through the same pipeline that judges ordinary messages, so image spam is assessed on its actual content.
This closes the obvious gap in any text-only defence and works automatically on every image, alongside perceptual image hashing that recognises a picture seen before even when it has been lightly re-encoded. Recognition covers eight languages — English, Russian, German, French, Spanish, Chinese, Japanese and Korean — which is where the honest limits sit: OCR is not perfect on heavily stylised fonts, distorted text or very low-resolution images, and a script outside that set will not be transcribed. It is a strong extra layer, not a guarantee that no text can hide in a picture, which is why it feeds the broader pipeline rather than deciding alone.
Frequently asked questions
Does scanning every image slow down my group or cost me anything?
Scanning runs automatically in the background as images arrive; members do not wait on it to keep chatting. It is part of the protection pipeline rather than a separate paid add-on. Perceptual image hashing also lets a picture that has already been seen and judged be recognised again quickly, so repeated spam images do not each need full re-analysis.
Will OCR misread normal images and flag innocent posts?
OCR only extracts text; it does not decide on its own. The extracted words and any decoded QR link are passed through the same detection pipeline as a normal message, so a harmless meme with text is judged on what that text actually says. As with all detection, borderline calls are logged and reviewable, and you can watch it in Monitoring Mode before enforcement is on.
Which languages can it actually read?
Recognition covers eight languages: English, Russian, German, French, Spanish, Chinese, Japanese and Korean. Text in a language outside that set will not be reliably transcribed, and heavily stylised fonts, distorted lettering or very low-resolution screenshots can defeat OCR — which is one reason it works as a layer feeding the wider pipeline rather than a standalone filter.
Can spammers beat OCR by distorting the text in an image?
Sometimes — deliberately warped or noisy text is exactly what OCR struggles with, and a determined spammer can degrade an image to hurt recognition. That is why OCR does not stand alone: QR decoding catches links hidden as codes, perceptual hashing catches reused images, and the AI pipeline judges whatever text is recovered, so beating one layer does not clear all of them.
Where teams use this
Real community types where this feature does the heavy lifting.
Protect your crypto group from scammers, fake airdrops, and phishing links — including scams hidden inside images.
Protect your fan base from scammers impersonating you. Build a safe space for your community to connect.
Shield your gaming server from trolls, spam bots, and toxic behavior. Let your members focus on playing.
Related Features
Explore more ways Telm can help your community