How to detect phone numbers and contact details in images

To find a phone number, a messaging handle, an e-mail address or a bank account written on a photo, check every uploaded image with an image moderation API that reads the text in pictures, and hold the images it flags for review. Detectivision's image check does this with its scam flag: one call per image, an answer with a confidence from 0 to 1, and no word list to keep.

Why scammers write contact details on photos

On a marketplace, a classifieds site, a rental site or a dating app, a scam usually needs one thing first: a way to talk to the buyer or the match somewhere else. Off the platform, its payment protection, its reports and its moderators no longer reach the conversation.

Most platforms already filter the text people type. A phone number in a listing title or a chat message is easy to catch. So the number moves into the picture:

• written on a sign held up in a profile photo;
• typed over a product photo, or added as a caption inside the image;
• spelled out in words or split with spaces, dots and emoji, so a pattern does not match it;
• given as a handle: "WhatsApp me", a Telegram name, a Snapchat or an Instagram username;
• given as a way to pay: an IBAN, a bank account or a crypto wallet address.

The same photos often carry the rest of the scam: a fake payment or bank screen that "proves" money was sent, a fake giveaway, a QR code that leads to a phishing page, or a promise of quick returns on an investment.

What to look for

Write down which of these your rules forbid, and where. A marketplace may forbid contact details in listing photos but allow a shop's address on its own page. A dating app may forbid them in profile photos and allow them in chat after a match.

How detection works

Every approach does two things: it reads the text in the picture, then decides whether that text is a way to reach someone.

1. Reading the text. Optical character recognition (OCR) turns the pixels into words. Text on a photo is harder than text on a page: it can be rotated, handwritten, stylised, small, or half hidden behind an object.
2. Deciding what the text is. Pattern matching looks for the shapes of phone numbers, e-mail addresses and account numbers. It is exact for the formats it knows and misses the ones it does not: a number in a country it does not cover, a number spelled out in words, a handle with no @ sign. A model that reads the picture as a whole, text and image together, can also judge a number written in words or a handle next to an app's logo.
3. Scoring and acting. The answer says what was found and how sure the check is. Your rules decide what happens next.

Some APIs give you only the first step: they return the words and where they are, and the rest is your code. Others do all three and answer with a category.

What to do when a photo is flagged

1. Hold, then decide. Keep a flagged photo out of view until a moderator has looked, or reject it at once if your rules forbid contact details everywhere. If your platform allows contact details in some places, treat the flag as a reason to look rather than to remove.
2. Look at the account, not only the photo. A new account with several flagged photos is not the same as a long-standing seller with one. Keep your user and listing ids with each check, so a moderator sees the history.
3. Tell the person why. Say which rule the photo broke and how to fix it, such as "remove the phone number and upload again", and give them a way to appeal.
4. Warn the other side. On a dating app or a marketplace chat, a short notice that moving to another app removes the platform's protection helps people who have already seen the photo.
5. Keep a record. Store the check's id, the flag, its confidence and the action taken. It answers appeals and shows where scams come from.

How Detectivision handles it

Send the photo as a file, its URL or base64 to POST /api/v1/moderation with "type": "image_moderation". By default every image flag of your account is checked in the same call, such as sexual, violence, hate and scam.

The scam flag covers fraud and contact details written on the picture. Detectivision's image docs list what counts: a phone number in digits or written out in words; a WhatsApp, Telegram, Snapchat or Instagram handle; an e-mail address; an IBAN or a crypto wallet; fake giveaways or prizes; phishing QR codes or links; fake payment or bank screens; and get-rich-quick or investment promises.

curl https://api.detectivision.ai/api/v1/moderation \
  -H "Api-Key: $DETECTIVISION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "image_moderation",
    "input": "https://example.com/listings/7731/1.jpg",
    "mode": "detect",
    "metadata": { "listingId": 7731, "sellerId": "s_204" }
  }'

{
  "id": 7120,
  "type": "image_moderation",
  "mode": "detect",
  "status": "completed",
  "flagged": true,
  "flags": [{ "flag": "scam", "confidence": 0.94 }],
  "checked": ["hate", "scam", "sexual", "violence"],
  "unchecked": [],
  "tookMs": 412,
  "usage": { "charged": 2, "remaining": 9998 },
  "metadata": { "listingId": 7731, "sellerId": "s_204" }
}

The answer says which flags the photo raised and how sure the check is. metadata comes back unchanged, so the result finds its listing and its seller. Values in the example are illustrative.

• Price. An image costs 1 credit when the answer comes by webhook ("async": true) and 2 when you wait for it. One image is checked for every flag for that price (Credits and billing).
• Imports and backlogs. Send listings already online with "async": true: each result comes to your webhook, signed, with your metadata (Async requests).
• Moderators. A moderator opens a check in the dashboard's Logs and takes one of your own actions, such as Hold or Ban. Your system gets it as a moderation.action webhook and carries it out (Metadata and actions).
• Images kept? No. An image is checked and not stored.

APIs that detect contact details in images

Each company's facts are from its own pages, as read on 5 November 2026. Detectivision is not affiliated with any of these companies. Prices and models change: read each page before you choose, and test with your own photos, including numbers written in words and handles without an @ sign.

Start with your own photos

Pick fifty photos your moderators removed for contact details and fifty they approved, and send all of them through the dashboard's Playground or the API. Count how many of the removed ones raise scam and how many of the approved ones do, then set your line on the confidence. The marketplace and dating app pages show which flags to check at each place on those platforms.